PP-StructureV3 Production Line User Guide¶

1. Introduction to PP-StructureV3 Production Line¶

Layout analysis is a technique used to extract structured information from document images. It is primarily used to convert complex document layouts into machine-readable data formats. This technology has broad applications in document management, information extraction, and data digitization. Layout analysis combines Optical Character Recognition (OCR), image processing, and machine learning algorithms to identify and extract text blocks, titles, paragraphs, images, tables, and other layout elements from documents. This process generally includes three main steps: layout analysis, element analysis, and data formatting. The final result is structured document data, which enhances the efficiency and accuracy of data processing. PP-StructureV3 improves upon the general layout analysis v1 production line by enhancing layout region detection, table recognition, and formula recognition. It also adds capabilities such as multi-column reading order recovery and result conversion to Markdown files. It performs excellently across various document types and can handle complex document data. This production line also provides flexible service deployment options, supporting invocation using multiple programming languages on various hardware. In addition, it offers secondary development capabilities, allowing you to train and fine-tune models on your own dataset and integrate the trained models seamlessly.

PP-StructureV3 includes the following six modules. Each module can be independently trained and inferred, and contains multiple models. Click the corresponding module for more documentation.

Layout Detection Module

General OCR Subline

Document Image Preprocessing Subline （Optional）

Table Recognition Subline （Optional）

Seal Recognition Subline （Optional）

Formula Recognition Subline （Optional）

Chart Parsing Module （Optional）

In this pipeline, you can choose the model to use based on the benchmark data below.

Document Image Orientation Classification Module :

Model Download Link Top-1 Acc (%) GPU Inference Time (ms)
[Standard Mode / High Performance Mode] CPU Inference Time (ms)
[Standard Mode / High Performance Mode] Model Size (M) Description

PP-LCNet_x1_0_doc_ori Inference Model/Pretrained Model 99.06 2.31 / 0.43 3.37 / 1.27 7 Document image classification model based on PP-LCNet_x1_0, supporting four categories: 0°, 90°, 180°, 270°

Text Image Rectification Module:

Text Image Rectification Module (Optional):

Model Download Link CER Model Size (M) Description

UVDoc Inference Model/Pretrained Model 0.179 30.3 M High-precision text image rectification model

Layout Detection Module Model:
* The layout detection model includes 20 common categories: document title, paragraph title, text, page number, abstract, table, references, footnotes, header, footer, algorithm, formula, formula number, image, table, seal, figure_table title, chart, and sidebar text and lists of references

Model Model Download Link mAP(0.5) (%) GPU Inference Time (ms)
[Normal Mode / High-Performance Mode] CPU Inference Time (ms)
[Normal Mode / High-Performance Mode] Model Storage Size (M) Introduction

PP-DocLayout_plus-L Inference Model/Training Model 83.2 34.6244 / 10.3945 510.57 / - 126.01 M A higher-precision layout area localization model trained on a self-built dataset containing Chinese and English papers, PPT, multi-layout magazines, contracts, books, exams, ancient books and research reports using RT-DETR-L

* The layout detection model includes 1 category: Block:

Model Model Download Link mAP(0.5) (%) GPU Inference Time (ms)
[Normal Mode / High-Performance Mode] CPU Inference Time (ms)
[Normal Mode / High-Performance Mode] Model Storage Size (M) Introduction

PP-DocBlockLayout Inference Model/Training Model 95.9 34.6244 / 10.3945 510.57 / - 123.92 M A layout block localization model trained on a self-built dataset containing Chinese and English papers, PPT, multi-layout magazines, contracts, books, exams, ancient books and research reports using RT-DETR-L

* The layout detection model includes 23 common categories: document title, paragraph title, text, page number, abstract, table of contents, references, footnotes, header, footer, algorithm, formula, formula number, image, figure caption, table, table caption, seal, figure title, figure, header image, footer image, and sidebar text

Model Download Link mAP(0.5) (%) GPU Inference Time (ms)
[Standard Mode / High Performance Mode] CPU Inference Time (ms)
[Standard Mode / High Performance Mode] Model Size (M) Description

PP-DocLayout-L Inference Model/Pretrained Model 90.4 34.6244 / 10.3945 510.57 / - 123.76 M A high-precision layout area localization model trained on a self-built dataset containing Chinese and English papers, magazines, contracts, books, exams, and research reports using RT-DETR-L.

PP-DocLayout-M Inference Model/Pretrained Model 75.2 13.3259 / 4.8685 44.0680 / 44.0680 22.578 A layout area localization model with balanced precision and efficiency, trained on a self-built dataset containing Chinese and English papers, magazines, contracts, books, exams, and research reports using PicoDet-L.

PP-DocLayout-S Inference Model/Pretrained Model 70.9 8.3008 / 2.3794 10.0623 / 9.9296 4.834 A high-efficiency layout area localization model trained on a self-built dataset containing Chinese and English papers, magazines, contracts, books, exams, and research reports using PicoDet-S.

> ❗ The above list includes the 4 core models that are key supported by the text recognition module. The module actually supports a total of 12 full models, including several predefined models with different categories. The complete model list is as follows:
👉 Details of Model List
* Table Layout Detection Model

Model Model Download Link mAP(0.5) (%) GPU Inference Time (ms)
[Normal Mode / High-Performance Mode] CPU Inference Time (ms)
[Normal Mode / High-Performance Mode] Model Storage Size (M) Introduction

PicoDet_layout_1x_table Inference Model/Training Model 97.5 8.02 / 3.09 23.70 / 20.41 7.4 M A high-efficiency layout area localization model trained on a self-built dataset using PicoDet-1x, capable of detecting table regions.

* 3-Class Layout Detection Model, including Table, Image, and Stamp

Model Model Download Link mAP(0.5) (%) GPU Inference Time (ms)
[Normal Mode / High-Performance Mode] CPU Inference Time (ms)
[Normal Mode / High-Performance Mode] Model Storage Size (M) Introduction

PicoDet-S_layout_3cls Inference Model/Training Model 88.2 8.99 / 2.22 16.11 / 8.73 4.8 A high-efficiency layout area localization model trained on a self-built dataset of Chinese and English papers, magazines, and research reports using PicoDet-S.

PicoDet-L_layout_3cls Inference Model/Training Model 89.0 13.05 / 4.50 41.30 / 41.30 22.6 A balanced efficiency and precision layout area localization model trained on a self-built dataset of Chinese and English papers, magazines, and research reports using PicoDet-L.

Table Classification Module Models:

RT-DETR-H_layout_3cls Inference Model/Training Model 95.8 114.93 / 27.71 947.56 / 947.56 470.1 A high-precision layout area localization model trained on a self-built dataset of Chinese and English papers, magazines, and research reports using RT-DETR-H.

* 5-Class English Document Area Detection Model, including Text, Title, Table, Image, and List

Model Model Download Link mAP(0.5) (%) GPU Inference Time (ms)
[Normal Mode / High-Performance Mode] CPU Inference Time (ms)
[Normal Mode / High-Performance Mode] Model Storage Size (M) Introduction

PicoDet_layout_1x Inference Model/Training Model 97.8 9.03 / 3.10 25.82 / 20.70 7.4 A high-efficiency English document layout area localization model trained on the PubLayNet dataset using PicoDet-1x.

* 17-Class Area Detection Model, including 17 common layout categories: Paragraph Title, Image, Text, Number, Abstract, Content, Figure Caption, Formula, Table, Table Caption, References, Document Title, Footnote, Header, Algorithm, Footer, and Stamp

Model Model Download Link mAP(0.5) (%) GPU Inference Time (ms)
[Normal Mode / High-Performance Mode] CPU Inference Time (ms)
[Normal Mode / High-Performance Mode] Model Storage Size (M) Introduction

PicoDet-S_layout_17cls Inference Model/Training Model 87.4 9.11 / 2.12 15.42 / 9.12 4.8 A high-efficiency layout area localization model trained on a self-built dataset of Chinese and English papers, magazines, and research reports using PicoDet-S.

PicoDet-L_layout_17cls Inference Model/Training Model 89.0 13.50 / 4.69 43.32 / 43.32 22.6 A balanced efficiency and precision layout area localization model trained on a self-built dataset of Chinese and English papers, magazines, and research reports using PicoDet-L.

RT-DETR-H_layout_17cls Inference Model/Training Model 98.3 115.29 / 104.09 995.27 / 995.27 470.2 A high-precision layout area localization model trained on a self-built dataset of Chinese and English papers, magazines, and research reports using RT-DETR-H.

Table Structure Recognition Module (Optional):

Model Download Link mAP (%) GPU Inference Time (ms)
[Standard Mode / High Performance Mode] CPU Inference Time (ms)
[Standard Mode / High Performance Mode] Model Size (M) Description

RT-DETR-L_wired_table_cell_det Inference Model/Pretrained Model 82.7 35.00 / 10.45 495.51 / 495.51 124M RT-DETR is the first real-time end-to-end object detection model. Based on RT-DETR-L, the PaddlePaddle Vision Team pre-trained the model on a custom table cell detection dataset, achieving good performance for both wired and wireless tables.

RT-DETR-L_wireless_table_cell_det Inference Model/Pretrained Model

Text Detection Module (Required):

Model Download Link Detection Hmean (%) GPU Inference Time (ms)
[Standard Mode / High Performance Mode] CPU Inference Time (ms)
[Standard Mode / High Performance Mode] Model Size (M) Description

PP-OCRv4_server_det Inference Model/Training Model 82.56 83.34 / 80.91 442.58 / 442.58 109 The server-side text detection model of PP-OCRv4, with higher accuracy, suitable for deployment on high-performance servers.

PP-OCRv4_mobile_det Inference Model/Training Model 77.35 8.79 / 3.13 51.00 / 28.58 4.7 The mobile text detection model of PP-OCRv4, with higher efficiency, suitable for deployment on edge devices.

PP-OCRv3_mobile_det Inference Model/Training Model 78.68 8.44 / 2.91 27.87 / 27.87 2.1 The mobile text detection model of PP-OCRv3, with higher efficiency, suitable for deployment on edge devices.

PP-OCRv3_server_det Inference Model/Training Model 80.11 65.41 / 13.67 305.07 / 305.07 102.1 The server-side text detection model of PP-OCRv3, with higher accuracy, suitable for deployment on high-performance servers.

Text Recognition Module Model (Required):

👉Full Model List
* PP-OCRv5 Multi-Scenario Models

Model Download Link Chinese Avg Accuracy (%) English Avg Accuracy (%) Traditional Chinese Avg Accuracy (%) Japanese Avg Accuracy (%) GPU Inference Time (ms)
[Standard Mode / High Performance Mode] CPU Inference Time (ms)
[Standard Mode / High Performance Mode] Model Size (M) Description

PP-OCRv5_server_rec Inference Model/Pretrained Model 86.38 64.70 93.29 60.35 - - 205 M PP-OCRv5_server_rec is a new-generation text recognition model. It efficiently and accurately supports four major languages: Simplified Chinese, Traditional Chinese, English, and Japanese, as well as handwriting, vertical text, pinyin, and rare characters, offering robust and efficient support for document understanding.

PP-OCRv5_mobile_rec Inference Model/Pretrained Model 81.29 66.00 83.55 54.65 - - 136 M PP-OCRv5_mobile_rec is a new-generation text recognition model. It efficiently and accurately supports four major languages: Simplified Chinese, Traditional Chinese, English, and Japanese, as well as handwriting, vertical text, pinyin, and rare characters, offering robust and efficient support for document understanding.

* Chinese Recognition Models

Model Download Link Avg Accuracy (%) GPU Inference Time (ms)
[Standard Mode / High Performance Mode] CPU Inference Time (ms)
[Standard Mode / High Performance Mode] Model Size (M) Description

PP-OCRv4_server_rec_doc Inference Model/Pretrained Model 86.58 6.65 / 2.38 32.92 / 32.92 91 M Based on PP-OCRv4_server_rec, trained on additional Chinese documents and PP-OCR mixed data. It supports over 15,000 characters including Traditional Chinese, Japanese, and special symbols, enhancing both document-specific and general text recognition accuracy.

PP-OCRv4_mobile_rec Inference Model/Pretrained Model 83.28 4.82 / 1.20 16.74 / 4.64 11 M Lightweight model of PP-OCRv4 with high inference efficiency, suitable for deployment on various edge devices.

PP-OCRv4_server_rec Inference Model/Pretrained Model 85.19 6.58 / 2.43 33.17 / 33.17 87 M Server-side model of PP-OCRv4 with high recognition accuracy, suitable for deployment on various servers.

PP-OCRv3_mobile_rec Inference Model/Pretrained Model 75.43 5.87 / 1.19 9.07 / 4.28 11 M Lightweight model of PP-OCRv3 with high inference efficiency, suitable for deployment on various edge devices.

Model Download Link Avg Accuracy (%) GPU Inference Time (ms)
[Standard Mode / High Performance Mode] CPU Inference Time (ms)
[Standard Mode / High Performance Mode] Model Size (M) Description

ch_SVTRv2_rec Inference Model/Pretrained Model 68.81 8.08 / 2.74 50.17 / 42.50 73.9 M SVTRv2 is a server-side recognition model developed by the OpenOCR team at Fudan University’s FVL Lab. It won first place in the OCR End-to-End Recognition task of the PaddleOCR Model Challenge, improving end-to-end accuracy on Benchmark A by 6% compared to PP-OCRv4.

Model Download Link Avg Accuracy (%) GPU Inference Time (ms)
[Standard Mode / High Performance Mode] CPU Inference Time (ms)
[Standard Mode / High Performance Mode] Model Size (M) Description

ch_RepSVTR_rec Inference Model/Pretrained Model 65.07 5.93 / 1.62 20.73 / 7.32 22.1 M RepSVTR is a mobile text recognition model based on SVTRv2. It won first place in the OCR End-to-End Recognition task of the PaddleOCR Model Challenge, improving accuracy on Benchmark B by 2.5% over PP-OCRv4 with comparable inference speed.

* English Recognition Models

Model Download Link Avg Accuracy (%) GPU Inference Time (ms)
[Standard Mode / High Performance Mode] CPU Inference Time (ms)
[Standard Mode / High Performance Mode] Model Size (M) Description

en_PP-OCRv4_mobile_rec Inference Model/Pretrained Model 70.39 4.81 / 0.75 16.10 / 5.31 6.8 M Ultra-lightweight English recognition model trained on PP-OCRv4, supporting English and number recognition.

en_PP-OCRv3_mobile_rec Inference Model/Pretrained Model 70.69 5.44 / 0.75 8.65 / 5.57 7.8 M Ultra-lightweight English recognition model trained on PP-OCRv3, supporting English and number recognition.

* Multilingual Recognition Models

Model Model Download Link Recognition Avg Accuracy(%) GPU Inference Time (ms)
[Normal / High Performance] CPU Inference Time (ms)
[Normal / High Performance] Model Size (M) Description

korean_PP-OCRv3_mobile_rec Inference Model/Pretrained Model 60.21 5.40 / 0.97 9.11 / 4.05 8.6 M An ultra-lightweight Korean text recognition model trained based on PP-OCRv3, supporting Korean and digits recognition

japan_PP-OCRv3_mobile_rec Inference Model/Pretrained Model 45.69 5.70 / 1.02 8.48 / 4.07 8.8 M An ultra-lightweight Japanese text recognition model trained based on PP-OCRv3, supporting Japanese and digits recognition

chinese_cht_PP-OCRv3_mobile_rec Inference Model/Pretrained Model 82.06 5.90 / 1.28 9.28 / 4.34 9.7 M An ultra-lightweight Traditional Chinese text recognition model trained based on PP-OCRv3, supporting Traditional Chinese and digits recognition

te_PP-OCRv3_mobile_rec Inference Model/Pretrained Model 95.88 5.42 / 0.82 8.10 / 6.91 7.8 M An ultra-lightweight Telugu text recognition model trained based on PP-OCRv3, supporting Telugu and digits recognition

ka_PP-OCRv3_mobile_rec Inference Model/Pretrained Model 96.96 5.25 / 0.79 9.09 / 3.86 8.0 M An ultra-lightweight Kannada text recognition model trained based on PP-OCRv3, supporting Kannada and digits recognition

ta_PP-OCRv3_mobile_rec Inference Model/Pretrained Model 76.83 5.23 / 0.75 10.13 / 4.30 8.0 M An ultra-lightweight Tamil text recognition model trained based on PP-OCRv3, supporting Tamil and digits recognition

latin_PP-OCRv3_mobile_rec Inference Model/Pretrained Model 76.93 5.20 / 0.79 8.83 / 7.15 7.8 M An ultra-lightweight Latin text recognition model trained based on PP-OCRv3, supporting Latin and digits recognition

arabic_PP-OCRv3_mobile_rec Inference Model/Pretrained Model 73.55 5.35 / 0.79 8.80 / 4.56 7.8 M An ultra-lightweight Arabic script recognition model trained based on PP-OCRv3, supporting Arabic script and digits recognition

cyrillic_PP-OCRv3_mobile_rec Inference Model/Pretrained Model 94.28 5.23 / 0.76 8.89 / 3.88 7.9 M An ultra-lightweight Cyrillic script recognition model trained based on PP-OCRv3, supporting Cyrillic script and digits recognition

devanagari_PP-OCRv3_mobile_rec Inference Model/Pretrained Model 96.44 5.22 / 0.79 8.56 / 4.06 7.9 M An ultra-lightweight Devanagari script recognition model trained based on PP-OCRv3, supporting Devanagari script and digits recognition

Text Line Orientation Classification Module (Optional):

Model Model Download Link Top-1 Acc (%) GPU Inference Time (ms)
[Normal / High Performance] CPU Inference Time (ms)
[Normal / High Performance] Model Size (M) Description

PP-LCNet_x0_25_textline_ori Inference Model/Pretrained Model 95.54 - - 0.32 A text line classification model based on PP-LCNet_x0_25, containing two categories: 0 degrees and 180 degrees

Formula Recognition Module (Optional):

Model Model Download Link En-BLEU(%) Zh-BLEU(%) GPU Inference Time (ms)
[Normal Mode / High-Performance Mode] CPU Inference Time (ms)
[Normal Mode / High-Performance Mode] Model Storage Size (M) Introduction

UniMERNet Inference Model/Training Model 85.91 43.50 2266.96/- -/- 1.53 G UniMERNet is a formula recognition model developed by Shanghai AI Lab. It uses Donut Swin as the encoder and MBartDecoder as the decoder. The model is trained on a dataset of one million samples, including simple formulas, complex formulas, scanned formulas, and handwritten formulas, significantly improving the recognition accuracy of real-world formulas.
PP-FormulaNet-S Inference Model/Training Model 87.00 45.71 202.25/- -/- 224 M PP-FormulaNet is an advanced formula recognition model developed by the Baidu PaddlePaddle Vision Team. The PP-FormulaNet-S version uses PP-HGNetV2-B4 as its backbone network. Through parallel masking and model distillation techniques, it significantly improves inference speed while maintaining high recognition accuracy, making it suitable for applications requiring fast inference. The PP-FormulaNet-L version, on the other hand, uses Vary_VIT_B as its backbone network and is trained on a large-scale formula dataset, showing significant improvements in recognizing complex formulas compared to PP-FormulaNet-S.

PP-FormulaNet-L Inference Model/Training Model 90.36 45.78 1976.52/- -/- 695 M
PP-FormulaNet_plus-S Inference Model/Training Model 88.71 53.32 191.69/- -/- 248 M PP-FormulaNet_plus is an enhanced version of the formula recognition model developed by the Baidu PaddlePaddle Vision Team, building upon the original PP-FormulaNet. Compared to the original version, PP-FormulaNet_plus utilizes a more diverse formula dataset during training, including sources such as Chinese dissertations, professional books, textbooks, exam papers, and mathematics journals. This expansion significantly improves the model’s recognition capabilities. Among the models, PP-FormulaNet_plus-M and PP-FormulaNet_plus-L have added support for Chinese formulas and increased the maximum number of predicted tokens for formulas from 1,024 to 2,560, greatly enhancing the recognition performance for complex formulas. Meanwhile, the PP-FormulaNet_plus-S model focuses on improving the recognition of English formulas. With these improvements, the PP-FormulaNet_plus series models perform exceptionally well in handling complex and diverse formula recognition tasks.

PP-FormulaNet_plus-M Inference Model/Training Model 91.45 89.76 1301.56/- -/- 592 M

PP-FormulaNet_plus-L Inference Model/Training Model 92.22 90.64 1745.25/- -/- 698 M

LaTeX_OCR_rec Inference Model/Training Model 74.55 39.96 1244.61/- -/- 99 M LaTeX-OCR is a formula recognition algorithm based on an autoregressive large model. It uses Hybrid ViT as the backbone network and a transformer as the decoder, significantly improving the accuracy of formula recognition.

Seal Text Detection Module (Optional):

Model Model Download Link Detection Hmean (%) GPU Inference Time (ms)
[Normal / High Performance] CPU Inference Time (ms)
[Normal / High Performance] Model Size (M) Description

PP-OCRv4_server_seal_det Inference Model/Pretrained Model 98.21 74.75 / 67.72 382.55 / 382.55 109 Server-side seal text detection model based on PP-OCRv4, offering higher accuracy and suitable for deployment on high-performance servers

PP-OCRv4_mobile_seal_det Inference Model/Pretrained Model 96.47 7.82 / 3.09 48.28 / 23.97 4.6 Mobile-side seal text detection model based on PP-OCRv4, offering higher efficiency and suitable for edge-side deployment

Chart Parsing Model Module：

Model Model Download Link Model parameter size（B） Model Storage Size (GB) Model Score Description

PP-Chart2Table Inference Model 0.58 1.4 75.98 PP-Chart2Table is a self-developed multimodal model by the PaddlePaddle team, focusing on chart parsing, demonstrating outstanding performance in both Chinese and English chart parsing tasks. The team adopted a carefully designed data generation strategy, constructing a high-quality multimodal dataset of nearly 700,000 entries covering common chart types like pie charts, bar charts, stacked area charts, and various application scenarios. They also designed a two-stage training method, utilizing large model distillation to fully leverage massive unlabeled OOD data. In internal business tests in both Chinese and English scenarios, PP-Chart2Table not only achieved the SOTA level among models of the same parameter scale but also reached accuracy comparable to 7B parameter scale VLM models in critical scenarios.

Test Environment Description:

Performance Test Environment

Test Dataset：

Document Image Orientation Classification Module: A self-built dataset using PaddleX, covering multiple scenarios such as ID cards and documents, containing 1000 images.

Text Image Rectification Model: DocUNet

Layout Region Detection Model: A self-built layout detection dataset using PaddleOCR, containing 10,000 images of common document types such as Chinese and English papers, magazines, and research reports.

Table Structure Recognition Model: A self-built English table recognition dataset using PaddleX.

Text Detection Model: A self-built Chinese dataset using PaddleOCR, covering multiple scenarios such as street scenes, web images, documents, and handwriting, with 500 images for detection.

Chinese Recognition Model: A self-built Chinese dataset using PaddleOCR, covering multiple scenarios such as street scenes, web images, documents, and handwriting, with 11,000 images for text recognition.

ch_SVTRv2_rec: Evaluation set A for "OCR End-to-End Recognition Task" in the PaddleOCR Algorithm Model Challenge

ch_RepSVTR_rec: Evaluation set B for "OCR End-to-End Recognition Task" in the PaddleOCR Algorithm Model Challenge.

English Recognition Model: A self-built English dataset using PaddleX.

Multilingual Recognition Model: A self-built multilingual dataset using PaddleX.

Text Line Orientation Classification Model: A self-built dataset using PaddleX, covering various scenarios such as ID cards and documents, containing 1000 images.

Seal Text Detection Model: A self-built dataset using PaddleX, containing 500 images of circular seal textures.

Hardware Configuration：

GPU: NVIDIA Tesla T4

CPU: Intel Xeon Gold 6271C @ 2.60GHz

Other Environments: Ubuntu 20.04 / cuDNN 8.6 / TensorRT 8.5.2.2

Inference Mode Description

Mode GPU Configuration CPU Configuration Acceleration Technology Combination

Normal Mode FP32 Precision / No TRT Acceleration FP32 Precision / 8 Threads PaddleInference

High-Performance Mode Optimal combination of pre-selected precision types and acceleration strategies FP32 Precision / 8 Threads Pre-selected optimal backend (Paddle/OpenVINO/TRT, etc.)

2. Quick Start¶

All the model pipelines provided by PaddleX can be quickly experienced. You can use the command line or Python on your local machine to experience the effect of the PP-StructureV3 pipeline.

Before using the PP-StructureV3 pipeline locally, please ensure that you have completed the installation of the PaddleX wheel package according to the PaddleOCR Local Installation Guide. If you wish to selectively install dependencies, please refer to the relevant instructions in the installation guide. The dependency group corresponding to this pipeline is ocr.

When performing GPU inference, the default configuration may use more than 16 GB of VRAM. Please ensure that your GPU has sufficient memory. To reduce VRAM usage, you can modify the configuration file as described below to disable unnecessary features.

2.1 Experiencing via Command Line¶

You can quickly experience the PP-StructureV3 pipeline with a single command.

paddleocr pp_structurev3 -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/pp_structure_v3_demo.png paddleocr pp_structurev3 -i ./pp_structure_v3_demo.png --use_doc_orientation_classify True paddleocr pp_structurev3 -i ./pp_structure_v3_demo.png --use_doc_unwarping True paddleocr pp_structurev3 -i ./pp_structure_v3_demo.png --use_textline_orientation False paddleocr pp_structurev3 -i ./pp_structure_v3_demo.png --device gpu

The parameter description can be found in 2.2 Python Script Integration. Supports specifying multiple devices simultaneously for parallel inference. For details, please refer to Pipeline Parallel Inference.

After running, the result will be printed to the terminal, as follows:

👉Click to Expand

{'res': {'input_path': 'pp_structure_v3_demo.png', 'model_settings': {'use_doc_preprocessor': False, 'use_general_ocr': True, 'use_seal_recognition': True, 'use_table_recognition': True, 'use_formula_recognition': True}, 'layout_det_res': {'input_path': None, 'page_index': None, 'boxes': [{'cls_id': 2, 'label': 'text', 'score': 0.9853514432907104, 'coordinate': [770.9531, 776.6814, 1122.6057, 1058.7322]}, {'cls_id': 1, 'label': 'image', 'score': 0.9848673939704895, 'coordinate': [775.7434, 202.27979, 1502.8113, 686.02136]}, {'cls_id': 2, 'label': 'text', 'score': 0.983731746673584, 'coordinate': [1152.3197, 1113.3275, 1503.3029, 1346.586]}, {'cls_id': 2, 'label': 'text', 'score': 0.9832221865653992, 'coordinate': [1152.5602, 801.431, 1503.8436, 986.3563]}, {'cls_id': 2, 'label': 'text', 'score': 0.9829439520835876, 'coordinate': [9.549545, 849.5713, 359.1173, 1058.7488]}, {'cls_id': 2, 'label': 'text', 'score': 0.9811657667160034, 'coordinate': [389.58298, 1137.2659, 740.66235, 1346.7488]}, {'cls_id': 2, 'label': 'text', 'score': 0.9775941371917725, 'coordinate': [9.1302185, 201.85, 359.0409, 339.05692]}, {'cls_id': 2, 'label': 'text', 'score': 0.9750366806983948, 'coordinate': [389.71454, 752.96924, 740.544, 889.92456]}, {'cls_id': 2, 'label': 'text', 'score': 0.9738152027130127, 'coordinate': [389.94565, 298.55988, 740.5585, 435.5124]}, {'cls_id': 2, 'label': 'text', 'score': 0.9737328290939331, 'coordinate': [771.50256, 1065.4697, 1122.2582, 1178.7324]}, {'cls_id': 2, 'label': 'text', 'score': 0.9728517532348633, 'coordinate': [1152.5154, 993.3312, 1503.2349, 1106.327]}, {'cls_id': 2, 'label': 'text', 'score': 0.9725610017776489, 'coordinate': [9.372787, 1185.823, 359.31738, 1298.7227]}, {'cls_id': 2, 'label': 'text', 'score': 0.9724331498146057, 'coordinate': [389.62848, 610.7389, 740.83234, 746.2377]}, {'cls_id': 2, 'label': 'text', 'score': 0.9720287322998047, 'coordinate': [389.29898, 897.0936, 741.41516, 1034.6616]}, {'cls_id': 2, 'label': 'text', 'score': 0.9713053703308105, 'coordinate': [10.323685, 1065.4663, 359.6786, 1178.8872]}, {'cls_id': 2, 'label': 'text', 'score': 0.9689728021621704, 'coordinate': [9.336395, 537.6609, 359.2901, 652.1881]}, {'cls_id': 2, 'label': 'text', 'score': 0.9684857130050659, 'coordinate': [10.7608185, 345.95068, 358.93616, 434.64087]}, {'cls_id': 2, 'label': 'text', 'score': 0.9681928753852844, 'coordinate': [9.674866, 658.89075, 359.56528, 770.4319]}, {'cls_id': 2, 'label': 'text', 'score': 0.9634978175163269, 'coordinate': [770.9464, 1281.1785, 1122.6522, 1346.7156]}, {'cls_id': 2, 'label': 'text', 'score': 0.96304851770401, 'coordinate': [390.0113, 201.28055, 740.1684, 291.53073]}, {'cls_id': 2, 'label': 'text', 'score': 0.962053120136261, 'coordinate': [391.21393, 1040.952, 740.5046, 1130.32]}, {'cls_id': 2, 'label': 'text', 'score': 0.9565253853797913, 'coordinate': [10.113251, 777.1482, 359.439, 842.437]}, {'cls_id': 2, 'label': 'text', 'score': 0.9497362375259399, 'coordinate': [390.31357, 537.86285, 740.47595, 603.9285]}, {'cls_id': 2, 'label': 'text', 'score': 0.9371236562728882, 'coordinate': [10.2034, 1305.9753, 359.5958, 1346.7295]}, {'cls_id': 0, 'label': 'paragraph_title', 'score': 0.9338151216506958, 'coordinate': [791.6062, 1200.8479, 1103.3257, 1259.9324]}, {'cls_id': 0, 'label': 'paragraph_title', 'score': 0.9326773285865784, 'coordinate': [408.0737, 457.37024, 718.9509, 516.63464]}, {'cls_id': 0, 'label': 'paragraph_title', 'score': 0.9274250864982605, 'coordinate': [29.448685, 456.6762, 340.99194, 515.6999]}, {'cls_id': 2, 'label': 'text', 'score': 0.8742568492889404, 'coordinate': [1154.7095, 777.3624, 1330.3086, 794.5853]}, {'cls_id': 2, 'label': 'text', 'score': 0.8442489504814148, 'coordinate': [586.49316, 160.15454, 927.468, 179.64203]}, {'cls_id': 11, 'label': 'doc_title', 'score': 0.8332607746124268, 'coordinate': [133.80017, 37.41908, 1380.8601, 124.1429]}, {'cls_id': 6, 'label': 'figure_title', 'score': 0.6770150661468506, 'coordinate': [812.1718, 705.1199, 1484.6973, 747.1692]}]}, 'overall_ocr_res': {'input_path': None, 'page_index': None, 'model_settings': {'use_doc_preprocessor': False, 'use_textline_orientation': False}, 'dt_polys': array([[[ 133, 35], ..., [ 133, 131]], ..., [[1154, 1323], ..., [1152, 1355]]], dtype=int16), 'text_det_params': {'limit_side_len': 960, 'limit_type': 'max', 'thresh': 0.3, 'box_thresh': 0.6, 'unclip_ratio': 2.0}, 'text_type': 'general', 'textline_orientation_angles': array([-1, ..., -1]), 'text_rec_score_thresh': 0.0, 'rec_texts': ['助力双方交往', '搭建友谊桥梁', '本报记者', '沈小晓', '任', '彦', '黄培昭', '身着中国传统民族服装的厄立特里亚青', '厄立特里亚高等教育与研究院合作建立，开', '年依次登台表演中国民族舞、现代舞、扇子舞', '设了中国语言课程和中国文化课程，注册学', '等,曼妙的舞姿赢得现场观众阵阵掌声。这', '生2万余人次。10余年来，厄特孔院已成为', '是日前厄立特里亚高等教育与研究院孔子学', '当地民众了解中国的一扇窗口。', '院(以下简称"厄特孔院")举办"喜迎新年"中国', '黄鸣飞表示，随着来学习中文的人日益', '歌舞比赛的场景。', '增多，阿斯马拉大学教学点已难以满足教学', '中国和厄立特里亚传统友谊深厚。近年', '需要。2024年4月，由中企蜀道集团所属四', '来,在高质量共建"一带一路"框架下，中厄两', '川路桥承建的孔院教学楼项目在阿斯马拉开', '国人文交流不断深化，互利合作的民意基础', '工建设，预计今年上半年峻工，建成后将为厄', '日益深厚。', '特孔院提供全新的办学场地。', '“学好中文，我们的', '“在中国学习的经历', '未来不是梦”', '让我看到更广阔的世界”', '“鲜花曾告诉我你怎样走过，大地知道你', '多年来，厄立特里亚广大赴华留学生和', '心中的每一个角落…"厄立特里亚阿斯马拉', '培训人员积极投身国家建设，成为助力该国', '大学综合楼二层，一阵优美的歌声在走廊里回', '发展的人才和厄中友好的见证者和推动者。', '响。循着熟悉的旋律轻轻推开一间教室的门，', '在厄立特里亚全国妇女联盟工作的约翰', '学生们正跟着老师学唱中文歌曲《同一首歌》。', '娜·特韦尔德·凯莱塔就是其中一位。她曾在', '这是厄特孔院阿斯马拉大学教学点的一', '中华女子学院攻读硕士学位，研究方向是女', '节中文歌曲课。为了让学生们更好地理解歌', '性领导力与社会发展。其间，她实地走访中国', '词大意，老师尤斯拉·穆罕默德萨尔·侯赛因逐', '多个地区，获得了观察中国社会发展的第一', '在厄立特里亚不久前举办的第六届中国风筝文化节上，当地小学生体验风筝制作。', '字翻译和解释歌词。随着伴奏声响起，学生们', '手资料。', '中国驻厄立特里亚大使馆供图', '边唱边随着节拍摇动身体，现场气氛热烈。', '谈起在中国求学的经历，约翰娜记忆犹', '“这是中文歌曲初级班，共有32人。学', '新："中国的发展在当今世界是独一无二的。', '“不管远近都是客人，请不用客气；相约', '瓦的北红海省博物馆。', '生大部分来自首都阿斯马拉的中小学，年龄', '沿着中国特色社会主义道路坚定前行，中国', '好了在一起我们欢迎你"在一场中厄青', '博物馆二层陈列着一个发掘自阿杜利', '最小的仅有6岁。”尤斯拉告诉记者。', '创造了发展奇迹，这一切都离不开中国共产党', '年联谊活动上，四川路桥中方员工同当地大', '斯古城的中国古代陶制酒器，罐身上写着', '尤斯拉今年23岁，是厄立特里亚一所公立', '的领导。中国的发展经验值得许多国家学习', '学生合唱《北京欢迎你》。厄立特里亚技术学', '“万""和""禅"“山"等汉字。“这件文物证', '学校的艺术老师。她12岁开始在厄特孔院学', '借鉴。”', '院计算机科学与工程专业学生鲁夫塔·谢拉', '明,很早以前我们就通过海上丝绸之路进行', '习中文,在2017年第十届"汉语桥"世界中学生', '正在西南大学学习的厄立特里亚博士生', '是其中一名演唱者，她很早便在孔院学习中', '贸易往来与文化交流。这也是厄立特里亚', '中文比赛中获得厄立特里亚赛区第一名，并和', '穆卢盖塔·泽穆伊对中国怀有深厚感情。8', '文，一直在为去中国留学作准备。"这句歌词', '与中国友好交往历史的有力证明。"北红海', '同伴代表厄立特里亚前往中国参加决赛，获得', '年前，在北京师范大学获得硕士学位后，穆卢', '是我们两国人民友谊的生动写照。无论是投', '省博物馆研究与文献部负责人伊萨亚斯·特', '团体优胜奖。2022年起，尤斯拉开始在厄特孔', '盖塔在社交媒体上写下这样一段话："这是我', '身于厄立特里亚基础设施建设的中企员工，', '斯法兹吉说。', '院兼职教授中文歌曲，每周末两个课时。“中国', '人生的重要一步，自此我拥有了一双坚固的', '还是在中国留学的厄立特里亚学子，两国人', '厄立特里亚国家博物馆考古学和人类学', '文化博大精深，我希望我的学生们能够通过中', '鞋子，赋予我穿越荆棘的力量。”', '民携手努力，必将推动两国关系不断向前发', '研究员菲尔蒙·特韦尔德十分喜爱中国文', '文歌曲更好地理解中国文化。"她说。', '穆卢盖塔密切关注中国在经济、科技、教', '展。"鲁夫塔说。', '化。他表示：“学习彼此的语言和文化，将帮', '“姐姐，你想去中国吗？""非常想！我想', '育等领域的发展，中国在科研等方面的实力', '厄立特里亚高等教育委员会主任助理萨', '助厄中两国人民更好地理解彼此，助力双方', '去看故宫、爬长城。"尤斯拉的学生中有一对', '与日俱增。在中国学习的经历让我看到更广', '马瑞表示："每年我们都会组织学生到中国访', '交往，搭建友谊桥梁。"', '能歌善舞的姐妹，姐姐露娅今年15岁，妹妹', '阔的世界，从中受益匪浅。', '问学习，目前有超过5000名厄立特里亚学生', '厄立特里亚国家博物馆馆长塔吉丁·努', '莉娅14岁，两人都已在厄特孔院学习多年，', '23岁的莉迪亚·埃斯蒂法诺斯已在厄特', '在中国留学。学习中国的教育经验，有助于', '里达姆·优素福曾多次访问中国，对中华文明', '中文说得格外流利。', '孔院学习3年，在中国书法、中国画等方面表', '提升厄立特里亚的教育水平。”', '的传承与创新、现代化博物馆的建设与发展', '露娅对记者说："这些年来，怀着对中文', '现十分优秀，在2024年厄立特里亚赛区的', '“共同向世界展示非', '印象深刻。“中国博物馆不仅有许多保存完好', '和中国文化的热爱，我们姐妹俩始终相互鼓', '“汉语桥"比赛中获得一等奖。莉迪亚说："学', '的文物，还充分运用先进科技手段进行展示，', '励，一起学习。我们的中文一天比一天好，还', '习中国书法让我的内心变得安宁和纯粹。我', '洲和亚洲的灿烂文明”', '帮助人们更好理解中华文明。"塔吉丁说，"厄', '学会了中文歌和中国舞。我们一定要到中国', '也喜欢中国的服饰,希望未来能去中国学习，', '立特里亚与中国都拥有悠久的文明，始终相', '去。学好中文，我们的未来不是梦！"', '把中国不同民族元素融入服装设计中，创作', '从阿斯马拉出发，沿着蜿蜓曲折的盘山', '互理解、相互尊重。我希望未来与中国同行', '据厄特孔院中方院长黄鸣飞介绍，这所', '出更多精美作品，也把厄特文化分享给更多', '公路一路向东寻找丝路印迹。驱车两个小', '加强合作，共同向世界展示非洲和亚洲的灿', '孔院成立于2013年3月，由贵州财经大学和', '的中国朋友。”', '时，记者来到位于厄立特里亚港口城市马萨', '烂文明。”'], 'rec_scores': array([0.99943757, ..., 0.98181838]), 'rec_polys': array([[[ 133, 35], ..., [ 133, 131]], ..., [[1154, 1323], ..., [1152, 1355]]], dtype=int16), 'rec_boxes': array([[ 133, ..., 131], ..., [1152, ..., 1359]], dtype=int16)}, 'text_paragraphs_ocr_res': {'rec_polys': array([[[ 133, 35], ..., [ 133, 131]], ..., [[1154, 1323], ..., [1152, 1355]]], dtype=int16), 'rec_texts': ['助力双方交往', '搭建友谊桥梁', '本报记者', '沈小晓', '任', '彦', '黄培昭', '身着中国传统民族服装的厄立特里亚青', '厄立特里亚高等教育与研究院合作建立，开', '年依次登台表演中国民族舞、现代舞、扇子舞', '设了中国语言课程和中国文化课程，注册学', '等,曼妙的舞姿赢得现场观众阵阵掌声。这', '生2万余人次。10余年来，厄特孔院已成为', '是日前厄立特里亚高等教育与研究院孔子学', '当地民众了解中国的一扇窗口。', '院(以下简称"厄特孔院")举办"喜迎新年"中国', '黄鸣飞表示，随着来学习中文的人日益', '歌舞比赛的场景。', '增多，阿斯马拉大学教学点已难以满足教学', '中国和厄立特里亚传统友谊深厚。近年', '需要。2024年4月，由中企蜀道集团所属四', '来,在高质量共建"一带一路"框架下，中厄两', '川路桥承建的孔院教学楼项目在阿斯马拉开', '国人文交流不断深化，互利合作的民意基础', '工建设，预计今年上半年峻工，建成后将为厄', '日益深厚。', '特孔院提供全新的办学场地。', '“学好中文，我们的', '“在中国学习的经历', '未来不是梦”', '让我看到更广阔的世界”', '“鲜花曾告诉我你怎样走过，大地知道你', '多年来，厄立特里亚广大赴华留学生和', '心中的每一个角落…"厄立特里亚阿斯马拉', '培训人员积极投身国家建设，成为助力该国', '大学综合楼二层，一阵优美的歌声在走廊里回', '发展的人才和厄中友好的见证者和推动者。', '响。循着熟悉的旋律轻轻推开一间教室的门，', '在厄立特里亚全国妇女联盟工作的约翰', '学生们正跟着老师学唱中文歌曲《同一首歌》。', '娜·特韦尔德·凯莱塔就是其中一位。她曾在', '这是厄特孔院阿斯马拉大学教学点的一', '中华女子学院攻读硕士学位，研究方向是女', '节中文歌曲课。为了让学生们更好地理解歌', '性领导力与社会发展。其间，她实地走访中国', '词大意，老师尤斯拉·穆罕默德萨尔·侯赛因逐', '多个地区，获得了观察中国社会发展的第一', '在厄立特里亚不久前举办的第六届中国风筝文化节上，当地小学生体验风筝制作。', '字翻译和解释歌词。随着伴奏声响起，学生们', '手资料。', '中国驻厄立特里亚大使馆供图', '边唱边随着节拍摇动身体，现场气氛热烈。', '谈起在中国求学的经历，约翰娜记忆犹', '“这是中文歌曲初级班，共有32人。学', '新："中国的发展在当今世界是独一无二的。', '“不管远近都是客人，请不用客气；相约', '瓦的北红海省博物馆。', '生大部分来自首都阿斯马拉的中小学，年龄', '沿着中国特色社会主义道路坚定前行，中国', '好了在一起我们欢迎你"在一场中厄青', '博物馆二层陈列着一个发掘自阿杜利', '最小的仅有6岁。”尤斯拉告诉记者。', '创造了发展奇迹，这一切都离不开中国共产党', '年联谊活动上，四川路桥中方员工同当地大', '斯古城的中国古代陶制酒器，罐身上写着', '尤斯拉今年23岁，是厄立特里亚一所公立', '的领导。中国的发展经验值得许多国家学习', '学生合唱《北京欢迎你》。厄立特里亚技术学', '“万""和""禅"“山"等汉字。“这件文物证', '学校的艺术老师。她12岁开始在厄特孔院学', '借鉴。”', '院计算机科学与工程专业学生鲁夫塔·谢拉', '明,很早以前我们就通过海上丝绸之路进行', '习中文,在2017年第十届"汉语桥"世界中学生', '正在西南大学学习的厄立特里亚博士生', '是其中一名演唱者，她很早便在孔院学习中', '贸易往来与文化交流。这也是厄立特里亚', '中文比赛中获得厄立特里亚赛区第一名，并和', '穆卢盖塔·泽穆伊对中国怀有深厚感情。8', '文，一直在为去中国留学作准备。"这句歌词', '与中国友好交往历史的有力证明。"北红海', '同伴代表厄立特里亚前往中国参加决赛，获得', '年前，在北京师范大学获得硕士学位后，穆卢', '是我们两国人民友谊的生动写照。无论是投', '省博物馆研究与文献部负责人伊萨亚斯·特', '团体优胜奖。2022年起，尤斯拉开始在厄特孔', '盖塔在社交媒体上写下这样一段话："这是我', '身于厄立特里亚基础设施建设的中企员工，', '斯法兹吉说。', '院兼职教授中文歌曲，每周末两个课时。“中国', '人生的重要一步，自此我拥有了一双坚固的', '还是在中国留学的厄立特里亚学子，两国人', '厄立特里亚国家博物馆考古学和人类学', '文化博大精深，我希望我的学生们能够通过中', '鞋子，赋予我穿越荆棘的力量。”', '民携手努力，必将推动两国关系不断向前发', '研究员菲尔蒙·特韦尔德十分喜爱中国文', '文歌曲更好地理解中国文化。"她说。', '穆卢盖塔密切关注中国在经济、科技、教', '展。"鲁夫塔说。', '化。他表示：“学习彼此的语言和文化，将帮', '“姐姐，你想去中国吗？""非常想！我想', '育等领域的发展，中国在科研等方面的实力', '厄立特里亚高等教育委员会主任助理萨', '助厄中两国人民更好地理解彼此，助力双方', '去看故宫、爬长城。"尤斯拉的学生中有一对', '与日俱增。在中国学习的经历让我看到更广', '马瑞表示："每年我们都会组织学生到中国访', '交往，搭建友谊桥梁。"', '能歌善舞的姐妹，姐姐露娅今年15岁，妹妹', '阔的世界，从中受益匪浅。', '问学习，目前有超过5000名厄立特里亚学生', '厄立特里亚国家博物馆馆长塔吉丁·努', '莉娅14岁，两人都已在厄特孔院学习多年，', '23岁的莉迪亚·埃斯蒂法诺斯已在厄特', '在中国留学。学习中国的教育经验，有助于', '里达姆·优素福曾多次访问中国，对中华文明', '中文说得格外流利。', '孔院学习3年，在中国书法、中国画等方面表', '提升厄立特里亚的教育水平。”', '的传承与创新、现代化博物馆的建设与发展', '露娅对记者说："这些年来，怀着对中文', '现十分优秀，在2024年厄立特里亚赛区的', '“共同向世界展示非', '印象深刻。“中国博物馆不仅有许多保存完好', '和中国文化的热爱，我们姐妹俩始终相互鼓', '“汉语桥"比赛中获得一等奖。莉迪亚说："学', '的文物，还充分运用先进科技手段进行展示，', '励，一起学习。我们的中文一天比一天好，还', '习中国书法让我的内心变得安宁和纯粹。我', '洲和亚洲的灿烂文明”', '帮助人们更好理解中华文明。"塔吉丁说，"厄', '学会了中文歌和中国舞。我们一定要到中国', '也喜欢中国的服饰,希望未来能去中国学习，', '立特里亚与中国都拥有悠久的文明，始终相', '去。学好中文，我们的未来不是梦！"', '把中国不同民族元素融入服装设计中，创作', '从阿斯马拉出发，沿着蜿蜓曲折的盘山', '互理解、相互尊重。我希望未来与中国同行', '据厄特孔院中方院长黄鸣飞介绍，这所', '出更多精美作品，也把厄特文化分享给更多', '公路一路向东寻找丝路印迹。驱车两个小', '加强合作，共同向世界展示非洲和亚洲的灿', '孔院成立于2013年3月，由贵州财经大学和', '的中国朋友。”', '时，记者来到位于厄立特里亚港口城市马萨', '烂文明。”'], 'rec_scores': array([0.99943757, ..., 0.98181838]), 'rec_boxes': array([[ 133, ..., 131], ..., [1152, ..., 1359]], dtype=int16)}}}

The result parameter description can be found in the result interpretation in 2.2.2 Python Script Integration.

Note: Since the default model of the pipeline is relatively large, the inference speed may be slow. You can refer to the model list in Section 1 and replace it with a model that has faster inference speed.

2.2 Python Script Integration¶

Just a few lines of code can complete the quick inference of the pipeline. Taking the PP-StructureV3 pipeline as an example:

from paddlex import create_pipeline pipeline = create_pipeline(pipeline="PP-StructureV3") ## 2. Quick Start Before using the PP-StructureV3 pipeline locally, please make sure you have completed the installation of the wheel package according to the [installation guide](../installation.en.md). After installation, you can use it via command line or Python integration. ### 2.1 Command Line Usage Use a single command to quickly experience the PP-StructureV3 pipeline: ```bash paddleocr pp_structurev3 -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/pp_structure_v3_demo.png # Use --use_doc_orientation_classify to enable document orientation classification paddleocr pp_structurev3 -i ./pp_structure_v3_demo.png --use_doc_orientation_classify True # Use --use_doc_unwarping to enable document unwarping module paddleocr pp_structurev3 -i ./pp_structure_v3_demo.png --use_doc_unwarping True # Use --use_textline_orientation to enable text line orientation classification paddleocr pp_structurev3 -i ./pp_structure_v3_demo.png --use_textline_orientation False # Use --device to specify GPU for inference paddleocr pp_structurev3 -i ./pp_structure_v3_demo.png --device gpu

Command line supports more parameters. Click to expand for detailed parameter descriptions

Parameter Description Type Default

input Data to be predicted. Required. Supports multiple input types.

Python Var: e.g., numpy.ndarray representing image data

str: e.g., local path to image or PDF file: /root/data/img.jpg; URL, e.g., online image or PDF: example; local directory: directory containing images to predict, e.g., /root/data/ (currently, directories with PDFs are not supported; PDFs must be specified by file path)

List: list elements must be one of the above types, e.g., [numpy.ndarray, numpy.ndarray], ["/root/data/img1.jpg", "/root/data/img2.jpg"], ["/root/data1", "/root/data2"]

Python Var|str|list

save_path Path to save inference results. If set to None, results will not be saved locally. str None

layout_detection_model_name Name of the layout detection model. If set to None, the default model will be used. str None

layout_detection_model_dir Directory path of the layout detection model. If set to None, the official model will be downloaded. str None

layout_threshold Score threshold for the layout model.

float: any value between 0-1

dict: {0:0.1}, where key is class ID and value is the threshold for that class

None: if set to None, the default value is used, which is 0.5

float|dict None

layout_nms Whether to apply NMS post-processing for layout detection model. bool None

layout_unclip_ratio Unclip ratio for detected boxes in layout detection model.

float: any float > 0

Tuple[float,float]: separate ratios for width and height

dict: key is int (class ID), value is tuple, e.g., {0: (1.1, 2.0)} means class 0 boxes will be expanded 1.1x in width, 2.0x in height

None: if set to None, default is 1.0

float|Tuple[float,float]|dict None

layout_merge_bboxes_mode Merge mode for overlapping boxes in layout detection.

str: large, small, union, for keeping larger box, smaller box, or both

dict: key is int (class ID), value is str, e.g., {0: "large", 2: "small"}

None: if set to None, default is large

str|dict None

chart_recognition_model_name Name of the chart recognition model. If set to None, the default model will be used. str None

chart_recognition_model_dir Directory path of the chart recognition model. If set to None, the official model will be downloaded. str None

chart_recognition_batch_size Batch size for the chart recognition model. If set to None, the default batch size is 1. int None

region_detection_model_name Name of the region detection model. If set to None, the default model will be used. str None

region_detection_model_dir Directory path of the region detection model. If set to None, the official model will be downloaded. str None

doc_orientation_classify_model_name Name of the document orientation classification model. If set to None, the default model will be used. str None

doc_orientation_classify_model_dir Directory path of the document orientation classification model. If set to None, the official model will be downloaded. str None

doc_unwarping_model_name Name of the document unwarping model. If set to None, the default model will be used. str None

doc_unwarping_model_dir Directory path of the document unwarping model. If set to None, the official model will be downloaded. str None

text_detection_model_name Name of the text detection model. If set to None, the default model will be used. str None

text_detection_model_dir Directory path of the text detection model. If set to None, the official model will be downloaded. str None

text_det_limit_side_len Maximum side length limit for text detection.

int: any integer > 0;

None: if set to None, the default value will be 960;

int None

text_det_limit_type

str: supports min and max; min means ensuring the shortest side of the image is not less than det_limit_side_len, max means the longest side does not exceed limit_side_len

None: if set to None, the default value will be max.

str None

text_det_thresh Pixel threshold for detection. Pixels with scores above this value in the probability map are considered text.

float: any float > 0

None: if set to None, default is 0.3

float None

text_det_box_thresh Box threshold. A bounding box is considered text if the average score of pixels inside is greater than this value.

float: any float > 0

None: if set to None, default is 0.6

float None

text_det_unclip_ratio Expansion ratio for text detection. The higher the value, the larger the expansion area.

float: any float > 0

None: if set to None, default is 2.0

float None

textline_orientation_model_name Name of the text line orientation model. If set to None, the default model will be used. str None

textline_orientation_model_dir Directory of the text line orientation model. If set to None, the official model will be downloaded. str None

textline_orientation_batch_size Batch size for the text line orientation model. If set to None, default is 1. int None

text_recognition_model_name Name of the text recognition model. If set to None, the default model will be used. str None

text_recognition_model_dir Directory of the text recognition model. If set to None, the official model will be downloaded. str None

text_recognition_batch_size Batch size for text recognition. If set to None, default is 1. int None

text_rec_score_thresh Score threshold for text recognition. Only results above this value will be kept.

float: any float > 0

None: if set to None, default is 0.0 (no threshold)

float None

table_classification_model_name Name of the table classification model. If set to None, the default model will be used. str None

table_classification_model_dir Directory of the table classification model. If set to None, the official model will be downloaded. str None

wired_table_structure_recognition_model_name Name of the wired table structure recognition model. If set to None, the default model will be used. str None

wired_table_structure_recognition_model_dir Directory of the wired table structure recognition model. If set to None, the official model will be downloaded. str None

wireless_table_structure_recognition_model_name Name of the wireless table structure recognition model. If set to None, the default model will be used. str None

wireless_table_structure_recognition_model_dir Directory of the wireless table structure recognition model. If set to None, the official model will be downloaded. str None

wired_table_cells_detection_model_name Name of the wired table cell detection model. If set to None, the default model will be used. str None

wired_table_cells_detection_model_dir Directory of the wired table cell detection model. If set to None, the official model will be downloaded. str None

wireless_table_cells_detection_model_name Name of the wireless table cell detection model. If set to None, the default model will be used. str None

wireless_table_cells_detection_model_dir Directory of the wireless table cell detection model. If set to None, the official model will be downloaded. str None

seal_text_detection_model_name Name of the seal text detection model. If set to None, the default model will be used. str None

seal_text_detection_model_dir Directory of the seal text detection model. If set to None, the official model will be downloaded. str None

seal_det_limit_side_len Image side length limit for seal text detection.

int: any integer > 0;

None: if set to None, the default is 736;

int None

seal_det_limit_type Limit type for image side in seal text detection.

str: supports min and max; min ensures shortest side ≥ det_limit_side_len, max ensures longest side ≤ limit_side_len

None: if set to None, default is min;

str None

seal_det_thresh Pixel threshold. Pixels with scores above this value in the probability map are considered text.

float: any float > 0

None: if set to None, default is 0.2

float None

seal_det_box_thresh Box threshold. Boxes with average pixel scores above this value are considered text regions.

float: any float > 0

None: if set to None, default is 0.6

float None

seal_det_unclip_ratio Expansion ratio for seal text detection. Higher value means larger expansion area.

float: any float > 0

None: if set to None, default is 0.5

float None

seal_text_recognition_model_name Name of the seal text recognition model. If set to None, the default model will be used. str None

seal_text_recognition_model_dir Directory of the seal text recognition model. If set to None, the official model will be downloaded. str None

seal_text_recognition_batch_size Batch size for seal text recognition. If set to None, default is 1. int None

seal_rec_score_thresh Recognition score threshold. Text results above this value will be kept.

float: any float > 0

None: if set to None, default is 0.0 (no threshold)

float None

formula_recognition_model_name Name of the formula recognition model. If set to None, the default model will be used. str None

formula_recognition_model_dir Directory of the formula recognition model. If set to None, the official model will be downloaded. str None

formula_recognition_batch_size Batch size of the formula recognition model. If set to None, the default is 1. int None

use_doc_orientation_classify Whether to enable document orientation classification. If set to None, default is True. bool None

use_doc_unwarping Whether to enable document unwarping. If set to None, default is True. bool None

use_seal_recognition Whether to enable seal recognition subpipeline. If set to None, default is True. bool None

use_table_recognition Whether to enable table recognition subpipeline. If set to None, default is True. bool None

use_formula_recognition Whether to enable formula recognition subpipeline. If set to None, default is True. bool None

use_chart_recognition Whether to enable chart recognition model. If set to None, default is True. bool None

use_region_detection Whether to enable region detection submodule for document images. If set to None, default is True. bool None

device Device for inference. You can specify a device ID.

CPU: e.g., cpu

GPU: e.g., gpu:0 means GPU 0

NPU: e.g., npu:0 means NPU 0

XPU: e.g., xpu:0 means XPU 0

MLU: e.g., mlu:0 means MLU 0

DCU: e.g., dcu:0 means DCU 0

None: If set to None, GPU 0 will be used by default if available; otherwise, CPU will be used.

str None

enable_hpi Whether to enable high performance inference. bool False

use_tensorrt Whether to use TensorRT for inference acceleration. bool False

min_subgraph_size Minimum subgraph size for optimizing subgraph execution. int 3

precision Computation precision, e.g., fp32, fp16. str fp32

enable_mkldnn Whether to enable MKL-DNN. If set to None, enabled by default. bool None

cpu_threads Number of threads to use when inferring on CPU. int 8

paddlex_config Path to the PaddleX pipeline configuration file. str None

The inference result will be printed in the terminal. The default output of the PP-StructureV3 pipeline is as follows:

👉Click to expand

{'res': {'input_path': '/root/.paddlex/predict_input/pp_structure_v3_demo.png', 'page_index': None, 'model_settings': {'use_doc_preprocessor': True, 'use_seal_recognition': True, 'use_table_recognition': True, 'use_formula_recognition': True}, 'doc_preprocessor_res': {'input_path': None, 'page_index': None, 'model_settings': {'use_doc_orientation_classify': True, 'use_doc_unwarping': True}, 'angle': 0}, 'layout_det_res': {'input_path': None, 'page_index': None, 'boxes': [{'cls_id': 2, 'label': 'text', 'score': 0.9848763942718506, 'coordinate': [743.2788696289062, 777.3158569335938, 1115.24755859375, 1067.84228515625]}, {'cls_id': 2, 'label': 'text', 'score': 0.9827454686164856, 'coordinate': [1137.95556640625, 1127.66943359375, 1524, 1367.6356201171875]}, {'cls_id': 1, 'label': 'image', 'score': 0.9813530445098877, 'coordinate': [755.2349243164062, 184.64149475097656, 1523.7294921875, 684.6146392822266]}, {'cls_id': 2, 'label': 'text', 'score': 0.980336606502533, 'coordinate': [350.7603759765625, 1148.5648193359375, 706.8020629882812, 1367.00341796875]}, {'cls_id': 2, 'label': 'text', 'score': 0.9798877239227295, 'coordinate': [1147.3890380859375, 802.6549072265625, 1523.9051513671875, 994.9046630859375]}, {'cls_id': 2, 'label': 'text', 'score': 0.9724758863449097, 'coordinate': [741.2205810546875, 1074.2657470703125, 1110.120849609375, 1191.2010498046875]}, {'cls_id': 2, 'label': 'text', 'score': 0.9724437594413757, 'coordinate': [355.6563415527344, 899.6616821289062, 710.9073486328125, 1042.1270751953125]}, {'cls_id': 2, 'label': 'text', 'score': 0.9723313450813293, 'coordinate': [0, 181.92404174804688, 334.43384313583374, 330.294677734375]}, {'cls_id': 2, 'label': 'text', 'score': 0.9720360636711121, 'coordinate': [356.7376403808594, 753.35302734375, 714.37841796875, 892.6129760742188]}, {'cls_id': 2, 'label': 'text', 'score': 0.9711183905601501, 'coordinate': [1144.5242919921875, 1001.2548217773438, 1524, 1120.6578369140625]}, {'cls_id': 2, 'label': 'text', 'score': 0.9707457423210144, 'coordinate': [0, 849.873291015625, 325.0664693713188, 1067.2911376953125]}, {'cls_id': 2, 'label': 'text', 'score': 0.9700680375099182, 'coordinate': [363.04437255859375, 289.2635498046875, 719.1571655273438, 427.5818786621094]}, {'cls_id': 2, 'label': 'text', 'score': 0.9693533182144165, 'coordinate': [359.4466857910156, 606.0006103515625, 717.9885864257812, 746.55126953125]}, {'cls_id': 2, 'label': 'text', 'score': 0.9682930111885071, 'coordinate': [0.050221771001815796, 1073.1942138671875, 323.85799154639244, 1191.3121337890625]}, {'cls_id': 2, 'label': 'text', 'score': 0.9649553894996643, 'coordinate': [0.7939082384109497, 1198.5465087890625, 321.2581721544266, 1317.218017578125]}, {'cls_id': 2, 'label': 'text', 'score': 0.9644040465354919, 'coordinate': [0, 337.225830078125, 332.2462143301964, 428.298583984375]}, {'cls_id': 2, 'label': 'text', 'score': 0.9637495279312134, 'coordinate': [365.5925598144531, 188.2151336669922, 718.556640625, 283.7483215332031]}, {'cls_id': 2, 'label': 'text', 'score': 0.9603620767593384, 'coordinate': [355.30633544921875, 1048.5457763671875, 708.771484375, 1141.828369140625]}, {'cls_id': 2, 'label': 'text', 'score': 0.9508902430534363, 'coordinate': [361.0450744628906, 530.7780151367188, 719.6325073242188, 599.1027221679688]}, {'cls_id': 2, 'label': 'text', 'score': 0.9459834694862366, 'coordinate': [0.035085976123809814, 532.7417602539062, 330.5401824116707, 772.7175903320312]}, {'cls_id': 0, 'label': 'paragraph_title', 'score': 0.9400503635406494, 'coordinate': [760.1524658203125, 1214.560791015625, 1085.24853515625, 1274.7890625]}, {'cls_id': 2, 'label': 'text', 'score': 0.9341079592704773, 'coordinate': [1.025873064994812, 777.8804931640625, 326.99016749858856, 844.8532104492188]}, {'cls_id': 0, 'label': 'paragraph_title', 'score': 0.9259933233261108, 'coordinate': [0.11050379276275635, 450.3547058105469, 311.77746546268463, 510.5243835449219]}, {'cls_id': 0, 'label': 'paragraph_title', 'score': 0.9208691716194153, 'coordinate': [380.79510498046875, 447.859130859375, 698.1744384765625, 509.0489807128906]}, {'cls_id': 2, 'label': 'text', 'score': 0.8683002591133118, 'coordinate': [1149.1656494140625, 778.3809814453125, 1339.960205078125, 796.5060424804688]}, {'cls_id': 2, 'label': 'text', 'score': 0.8455104231834412, 'coordinate': [561.3448486328125, 140.87547302246094, 915.4432983398438, 162.76724243164062]}, {'cls_id': 11, 'label': 'doc_title', 'score': 0.735536515712738, 'coordinate': [76.71978759765625, 0, 1400.4561157226562, 98.32131713628769]}, {'cls_id': 6, 'label': 'figure_title', 'score': 0.7187536954879761, 'coordinate': [790.4249267578125, 704.4551391601562, 1509.9013671875, 747.6876831054688]}, {'cls_id': 2, 'label': 'text', 'score': 0.6218013167381287, 'coordinate': [737.427001953125, 1296.2047119140625, 1104.2994384765625, 1368]}]}, 'overall_ocr_res': {'input_path': None, 'page_index': None, 'model_settings': {'use_doc_preprocessor': False, 'use_textline_orientation': True}, 'dt_polys': array([[[ 77, 0], ..., [ 76, 98]], ..., [[1142, 1350], ..., [1142, 1367]]], dtype=int16), 'text_det_params': {'limit_side_len': 736, 'limit_type': 'min', 'thresh': 0.3, 'box_thresh': 0.6, 'unclip_ratio': 1.5}, 'text_type': 'general', 'textline_orientation_angles': array([0, ..., 0]), 'text_rec_score_thresh': 0.0, 'rec_texts': ['助力双方交往', '搭建友谊桥梁', '本报记者沈小晓任彦', '黄培照', '身着中国传统民族服装的厄立特里亚青', '厄立特里亚高等教育与研究院合作建立，开', '次登台表演中国民族舞、现代舞、扇子舞', '设了中国语言课程和中国文化课程，注册学', '曼妙的舞姿赢得现场观众阵阵掌声。这', '生2万余人次。10余年来，厄特孔院已成为', '日前厄立特里亚高等教育与研究院孔子学', '当地民众了解中国的一扇窗口。', '以下简称"厄特孔院")举办“喜迎新年"中国', '黄鸣飞表示，随着来学习中文的人日益', '舞比赛的场景。', '增多，阿斯马拉大学教学点已难以满足教学', '中国和厄立特里亚传统友谊深厚。近年', '需要。2024年4月，由中企蜀道集团所属四', '在高质量共建"一带一路"框架下，中厄两', '川路桥承建的孔院教学楼项目在阿斯马拉开', '人文交流不断深化，互利合作的民意基础', '工建设，预计今年上半年竣工，建成后将为厄', '益深厚。', '特孔院提供全新的办学场地。', '学好中文，我们的', '□', '在中国学习的经历', '未来不是梦”', '让我看到更广阔的世界”', '“鲜花曾告诉我你怎样走过，大地知道你', '多年来，厄立特里亚广大赴华留学生和', '中的每一个角落"厄立特里亚阿斯马拉', '培训人员积极投身国家建设，成为助力该国', '综合楼二层，一阵优美的歌声在走廊里回', '发展的人才和厄中友好的见证者和推动者。', '循着熟悉的旋律轻轻推开一间教室的门，', '在厄立特里亚全国妇女联盟工作的约翰', '们正跟着老师学唱中文歌曲《同一首歌》。', '娜·特韦尔德·凯莱塔就是其中一位。她曾在', '这是厄特孔院阿斯马拉大学教学点的一', '中华女子学院攻读硕士学位，研究方向是女', '中文歌曲课。为了让学生们更好地理解歌', '性领导力与社会发展。其间，她实地走访中国', '大意，老师尤斯拉·穆罕默德萨尔·侯赛因逐', '多个地区，获得了观察中国社会发展的第一', '在厄立特里亚不久前举办的第六届中国风筝文化节上，当地小学生体验风筝制作。', '译和解释歌词。随着伴奏声响起，学生们', '手资料。', '中国驻厄立特里亚大使馆供图', '昌边随着节拍摇动身体，现场气氛热烈。', '谈起在中国求学的经历，约翰娜记忆犹', '“这是中文歌曲初级班，共有32人。学', '新：“中国的发展在当今世界是独一无二的。', '“不管远近都是客人，请不用客气；相约', '瓦的北红海省博物馆。', '大部分来自首都阿斯马拉的中小学，年龄', '沿着中国特色社会主义道路坚定前行，中国', '好了在一起，我们欢迎你…"在一场中厄青', '博物馆二层陈列着一个发掘自阿杜禾', '小的仅有6岁。"尤斯拉告诉记者。', '创造了发展奇迹，这一切都离不开中国共产党', '年联谊活动上，四川路桥中方员工同当地大', '斯古城的中国古代陶制酒器，罐身上写', '尤斯拉今年23岁，是厄立特里亚一所公立', '的领导。中国的发展经验值得许多国家学习', '学生合唱《北京欢迎你》。厄立特里亚技术学', '“万”“和”“禅”“山"等汉字。“这件文物证', '交的艺术老师。她12岁开始在厄特孔院学', '借鉴。”', '院计算机科学与工程专业学生鲁夫塔·谢拉', '明，很早以前我们就通过海上丝绸之路进行', '中文，在2017年第十届“汉语桥"世界中学生', '正在西南大学学习的厄立特里亚博士生', '是其中一名演唱者，她很早便在孔院学习中', '贸易往来与文化交流。这也是厄立特里亚', '文比赛中获得厄立特里亚赛区第一名，并和', '穆卢盖塔·泽穆伊对中国怀有深厚感情。8', '文，一直在为去中国留学作准备。“这句歌词', '与中国友好交往历史的有力证明。”北红海', '半代表厄立特里亚前往中国参加决赛，获得', '年前，在北京师范大学获得硕士学位后，穆卢', '是我们两国人民友谊的生动写照。无论是投', '省博物馆研究与文献部负责人伊萨亚斯·特', '本优胜奖。2022年起，尤斯拉开始在厄特孔', '盖塔在社交媒体上写下这样一段话：“这是我', '身于厄立特里亚基础设施建设的中企员工，', '斯法兹吉说。', '兼职教授中文歌曲，每周末两个课时。“中国', '人生的重要一步，自此我拥有了一双坚固的', '还是在中国留学的厄立特里亚学子，两国人', '厄立特里亚国家博物馆考古学和人类学', '化博大精深，我希望我的学生们能够通过中', '鞋子，赋予我穿越荆棘的力量。”', '民携手努力，必将推动两国关系不断向前发', '研究员菲尔蒙·特韦尔德十分喜爱中国文', '软曲更好地理解中国文化。"她说。', '穆卢盖塔密切关注中国在经济、科技、教', '展。"鲁夫塔说。', '化。他表示：“学习彼此的语言和文化，将帮', '“姐姐，你想去中国吗？"“非常想！我想', '育等领域的发展，“中国在科研等方面的实力', '厄立特里亚高等教育委员会主任助理萨', '助厄中两国人民更好地理解彼此，助力双方', '看故宫、爬长城。"尤斯拉的学生中有一对', '与日俱增。在中国学习的经历让我看到更广', '马瑞表示：“每年我们都会组织学生到中国访', '交往，搭建友谊桥梁。”', '软善舞的姐妹，姐姐露娅今年15岁，妹妹', '阔的世界，从中受益匪浅。”', '问学习，目前有超过5000名厄立特里亚学生', '厄立特里亚国家博物馆馆长塔吉丁·', '亚14岁，两人都已在厄特孔院学习多年，', '23岁的莉迪亚·埃斯蒂法诺斯已在厄特', '在中国留学。学习中国的教育经验，有助于', '里达姆·优素福曾多次访问中国，对中华文明', '文说得格外流利。', '孔院学习3年，在中国书法、中国画等方面表', '提升厄立特里亚的教育水平。”', '的传承与创新、现代化博物馆的建设与发用', '露娅对记者说：“这些年来，怀着对中文', '现十分优秀，在2024年厄立特里亚赛区的', '印象深刻。“中国博物馆不仅有许多保存完好', '“共同向世界展示非', '中国文化的热爱，我们姐妹俩始终相互鼓', '“汉语桥”比赛中获得一等奖。莉迪亚说：“学', '的文物，还充分运用先进科技手段进行展示', '一起学习。我们的中文一天比一天好，还', '习中国书法让我的内心变得安宁和纯粹。我', '洲和亚洲的灿烂文明”', '帮助人们更好理解中华文明。"塔吉丁说，“', '了中文歌和中国舞。我们一定要到中国', '也喜欢中国的服饰，希望未来能去中国学习，', '立特里亚与中国都拥有悠久的文明，始终木', '学好中文，我们的未来不是梦！”', '把中国不同民族元素融入服装设计中，创作', '从阿斯马拉出发，沿着蜿蜒曲折的盘山', '互理解、相互尊重。我希望未来与中国同行', '据厄特孔院中方院长黄鸣飞介绍，这所', '出更多精美作品，也把厄特文化分享给更多', '公路一路向东寻找丝路印迹。驱车两个小', '加强合作，共同向世界展示非洲和亚洲的灿', '中贵州财经大学和', '的中国朋友。”', '时，记者来到位于厄立特里亚港口城市马萨', '烂文明。”'], 'rec_scores': array([0.99875408, ..., 0.98324996]), 'rec_polys': array([[[ 77, 0], ..., [ 76, 98]], ..., [[1142, 1350], ..., [1142, 1367]]], dtype=int16), 'rec_boxes': array([[ 76, ..., 103], ..., [1142, ..., 1367]], dtype=int16)}}}

For explanation of the result parameters, refer to 2.2 Python Script Integration.

Note: Due to the large size of the default model in the pipeline, the inference speed may be slow. You can refer to the model list in Section 1 to replace it with a faster model.

2.2 Python Script Integration¶

The command line method is for quick testing and visualization. In actual projects, you usually need to integrate the model via code. You can perform pipeline inference with just a few lines of code as shown below:

from paddleocr import PPStructureV3 pipeline = PPStructureV3() # ocr = PPStructureV3(use_doc_orientation_classify=True) # Use use_doc_orientation_classify to enable/disable document orientation classification model # ocr = PPStructureV3(use_doc_unwarping=True) # Use use_doc_unwarping to enable/disable document unwarping module # ocr = PPStructureV3(use_textline_orientation=True) # Use use_textline_orientation to enable/disable textline orientation classification model # ocr = PPStructureV3(device="gpu") # Use device to specify GPU for model inference output = pipeline.predict("./pp_structure_v3_demo.png") for res in output: res.print() ## Print the structured prediction output res.save_to_json(save_path="output") ## Save the current image's structured result in JSON format res.save_to_markdown(save_path="output") ## Save the current image's result in Markdown format

For PDF files, each page will be processed individually and generate a separate Markdown file. If you want to convert the entire PDF to a single Markdown file, use the following method:

from pathlib import Path from paddleocr import PPStructureV3 input_file = "./your_pdf_file.pdf" output_path = Path("./output") pipeline = PPStructureV3() output = pipeline.predict("./pp_structure_v3_demo.png") markdown_list = [] markdown_images = [] for res in output: md_info = res.markdown markdown_list.append(md_info) markdown_images.append(md_info.get("markdown_images", {})) markdown_texts = pipeline.concatenate_markdown_pages(markdown_list) mkd_file_path = output_path / f"{Path(input_file).stem}.md" mkd_file_path.parent.mkdir(parents=True, exist_ok=True) with open(mkd_file_path, "w", encoding="utf-8") as f: f.write(markdown_texts) for item in markdown_images: if item: for path, image in item.items(): file_path = output_path / path file_path.parent.mkdir(parents=True, exist_ok=True) image.save(file_path)

Note:

The default text recognition model used by PP-StructureV3 is a Chinese-English recognition model, which has limited accuracy for purely English texts. For English-only scenarios, you can set the text_recognition_model_name parameter to an English model such as en_PP-OCRv4_mobile_rec to achieve better recognition performance. For other languages, refer to the model list above and select the appropriate language recognition model for replacement.

In the example code, the parameters use_doc_orientation_classify, use_doc_unwarping, and use_textline_orientation are all set to False by default. These indicate that document orientation classification, document image unwarping, and textline orientation classification are disabled. You can manually set them to True if needed.

The above Python script performs the following steps:

(1) Instantiate PPStructureV3 to create the pipeline object. The parameter descriptions are as follows:

Parameter Description Type Default

layout_detection_model_name Name of the layout detection model. If set to None, the pipeline default model is used. str None

layout_detection_model_dir Directory path of the layout detection model. If set to None, the official model will be downloaded. str None

layout_threshold Score threshold for the layout model.

float: Any float between 0-1;

dict: {0:0.1} where the key is the class ID and the value is the threshold for that class;

None: If set to None, uses the pipeline default of 0.5;

float|dict None

layout_nms Whether to use NMS post-processing for the layout detection model. bool None

layout_unclip_ratio Expansion ratio for the bounding boxes from the layout detection model.

float: Any float greater than 0;

Tuple[float,float]: Expansion ratios in horizontal and vertical directions;

dict: A dictionary with int keys representing cls_id, and tuple values, e.g., {0: (1.1, 2.0)} means width is expanded 1.1× and height 2.0× for class 0 boxes;

None: If set to None, uses the pipeline default of 1.0;

float|Tuple[float,float]|dict None

layout_merge_bboxes_mode Filtering method for overlapping boxes in layout detection.

str: Options include large, small, and union to retain the larger box, smaller box, or both;

dict: A dictionary with int keys representing cls_id, and str values, e.g., {0: "large", 2: "small"} means using different modes for different classes;

None: If set to None, uses the pipeline default value large;

str|dict None

chart_recognition_model_name Name of the chart recognition model. If set to None, the pipeline default model is used. str None

chart_recognition_model_dir Directory path of the chart recognition model. If set to None, the official model will be downloaded. str None

chart_recognition_batch_size Batch size for the chart recognition model. If set to None, the default is 1. int None

region_detection_model_name Name of the region detection model for sub-modules in document layout. If set to None, the pipeline default model is used. str None

region_detection_model_dir Directory path of the region detection model. If set to None, the official model will be downloaded. str None

doc_orientation_classify_model_name Name of the document orientation classification model. If set to None, the pipeline default model is used. str None

doc_orientation_classify_model_dir Directory path of the document orientation classification model. If set to None, the official model will be downloaded. str None

doc_unwarping_model_name Name of the document unwarping model. If set to None, the pipeline default model is used. str None

doc_unwarping_model_dir Directory path of the document unwarping model. If set to None, the official model will be downloaded. str None

text_detection_model_name Name of the text detection model. If set to None, the pipeline default model is used. str None

text_detection_model_dir Directory path of the text detection model. If set to None, the official model will be downloaded. str None

text_det_limit_side_len Maximum side length limit for text detection.

int: Any integer greater than 0;

None: If set to None, uses the pipeline default of 960;

int None

text_det_limit_type

str: Supports min and max. min ensures the shortest side is no less than det_limit_side_len, while max ensures the longest side is no greater than limit_side_len;

None: If set to None, uses the pipeline default of max;

str None

text_det_thresh Pixel threshold for detection. Pixels in the output probability map with scores above this value are considered as text pixels.

float: Any float greater than 0;

None: If set to None, uses the pipeline default value of 0.3;

float None

text_det_box_thresh Bounding box threshold. If the average score of all pixels inside the box exceeds this threshold, it is considered a text region.

float: Any float greater than 0;

None: If set to None, uses the pipeline default value of 0.6;

float None

text_det_unclip_ratio Expansion ratio for text detection. The larger the value, the more the text region is expanded.

float: Any float greater than 0;

None: If set to None, uses the pipeline default value of 2.0;

float None

textline_orientation_model_name Name of the textline orientation model. If set to None, the pipeline default model is used. str None

textline_orientation_model_dir Directory path of the textline orientation model. If set to None, the official model will be downloaded. str None

textline_orientation_batch_size Batch size for the textline orientation model. If set to None, the default batch size is 1. int None

text_recognition_model_name Name of the text recognition model. If set to None, the pipeline default model is used. str None

text_recognition_model_dir Directory path of the text recognition model. If set to None, the official model will be downloaded. str None

text_recognition_batch_size Batch size for the text recognition model. If set to None, the default batch size is 1. int None

text_rec_score_thresh Score threshold for text recognition. Only results with scores above this threshold will be retained.

float: Any float greater than 0;

None: If set to None, uses the pipeline default of 0.0 (no threshold);

float None

table_classification_model_name Name of the table classification model. If set to None, the pipeline default model is used. str None

table_classification_model_dir Directory path of the table classification model. If set to None, the official model will be downloaded. str None

wired_table_structure_recognition_model_name Name of the wired table structure recognition model. If set to None, the pipeline default model is used. str None

wired_table_structure_recognition_model_dir Directory path of the wired table structure recognition model. If set to None, the official model will be downloaded. str None

wireless_table_structure_recognition_model_name Name of the wireless table structure recognition model. If set to None, the pipeline default model is used. str None

wireless_table_structure_recognition_model_dir Directory path of the wireless table structure recognition model. If set to None, the official model will be downloaded. str None

wired_table_cells_detection_model_name Name of the wired table cell detection model. If set to None, the pipeline default model is used. str None

wired_table_cells_detection_model_dir Directory path of the wired table cell detection model. If set to None, the official model will be downloaded. str None

wireless_table_cells_detection_model_name Name of the wireless table cell detection model. If set to None, the pipeline default model is used. str None

wireless_table_cells_detection_model_dir Directory path of the wireless table cell detection model. If set to None, the official model will be downloaded. str None

seal_text_detection_model_name Name of the seal text detection model. If set to None, the pipeline default model is used. str None

seal_text_detection_model_dir Directory path of the seal text detection model. If set to None, the official model will be downloaded. str None

seal_det_limit_side_len Image side length limit for seal text detection.

int: Any integer greater than 0;

None: If set to None, the default value is 736;

int None

seal_det_limit_type Limit type for seal text detection image side length.

str: Supports min and max. min ensures the shortest side is no less than det_limit_side_len, while max ensures the longest side is no greater than limit_side_len;

None: If set to None, the default value is min;

str None

seal_det_thresh Pixel threshold for detection. Pixels with scores greater than this value in the probability map are considered text pixels.

float: Any float greater than 0;

None: If set to None, the default value is 0.2;

float None

seal_det_box_thresh Bounding box threshold. If the average score of all pixels inside a detection box exceeds this threshold, it is considered a text region.

float: Any float greater than 0;

None: If set to None, the default value is 0.6;

float None

seal_det_unclip_ratio Expansion ratio for seal text detection. The larger the value, the larger the expanded area.

float: Any float greater than 0;

None: If set to None, the default value is 0.5;

float None

seal_text_recognition_model_name Name of the seal text recognition model. If set to None, the pipeline default model is used. str None

seal_text_recognition_model_dir Directory path of the seal text recognition model. If set to None, the official model will be downloaded. str None

seal_text_recognition_batch_size Batch size for the seal text recognition model. If set to None, the default value is 1. int None

seal_rec_score_thresh Score threshold for seal text recognition. Text results with scores above this threshold will be retained.

float: Any float greater than 0;

None: If set to None, the default value is 0.0 (no threshold);

float None

formula_recognition_model_name Name of the formula recognition model. If set to None, the pipeline default model is used. str None

formula_recognition_model_dir Directory path of the formula recognition model. If set to None, the official model will be downloaded. str None

formula_recognition_batch_size Batch size for the formula recognition model. If set to None, the default value is 1. int None

use_doc_orientation_classify Whether to enable the document orientation classification module. If set to None, the default value is True. bool None

use_doc_unwarping Whether to enable the document image unwarping module. If set to None, the default value is True. bool None

use_chart_recognition Whether to enable the chart recognition model. If set to None, the default value is True. bool None

use_region_detection Whether to enable the region detection model for document layout. If set to None, the default value is True. bool None

device Device used for inference. Supports specifying device ID.

CPU: e.g., cpu means using CPU for inference;

GPU: e.g., gpu:0 means using GPU 0;

NPU: e.g., npu:0 means using NPU 0;

XPU: e.g., xpu:0 means using XPU 0;

MLU: e.g., mlu:0 means using MLU 0;

DCU: e.g., dcu:0 means using DCU 0;

None: If set to None, GPU 0 will be used by default. If GPU is not available, CPU will be used;

str None

enable_hpi Whether to enable high-performance inference. bool False

use_tensorrt Whether to use TensorRT for accelerated inference. bool False

min_subgraph_size Minimum subgraph size used to optimize model subgraph computation. int 3

precision Computation precision, e.g., fp32, fp16. str fp32

enable_mkldnn Whether to enable MKL-DNN acceleration. If set to None, MKL-DNN is enabled by default. bool None

cpu_threads Number of threads used for inference on CPU. int 8

paddlex_config Path to the PaddleX pipeline configuration file. str None

(2) Call the predict() method of the PP-StructureV3 pipeline object for inference. This method returns a result list. The pipeline also provides a predict_iter() method. Both methods accept the same parameters and return the same type of results. The only difference is that predict_iter() returns a generator that allows incremental processing and retrieval of prediction results, which is useful for handling large datasets or saving memory. Choose the method that fits your needs. Below are the parameters of the predict() method:

Parameter Description Type Default

input Input data to be predicted. Required. Supports multiple types:

Python Var: Image data represented as numpy.ndarray

str: Local path to image or PDF file, e.g., /root/data/img.jpg; URL to image or PDF, e.g., example; directory containing image files, e.g., /root/data/ (directories with PDFs are not supported, use full file path for PDFs)

List: Elements can be any of the above types, e.g., [numpy.ndarray, numpy.ndarray], ["/root/data/img1.jpg", "/root/data/img2.jpg"], ["/root/data1", "/root/data2"]

Python Var|str|list

device Same as the parameter used during initialization. str None

use_doc_orientation_classify Whether to use document orientation classification during inference. bool None

use_doc_unwarping Whether to use document image unwarping during inference. bool None

use_textline_orientation Whether to use textline orientation classification during inference. bool None

use_seal_recognition Whether to use the seal recognition sub-pipeline during inference. bool None

use_table_recognition Whether to use the table recognition sub-pipeline during inference. bool None

use_formula_recognition Whether to use the formula recognition sub-pipeline during inference. bool None

layout_threshold Same as the parameter used during initialization. float|dict None

layout_nms Same as the parameter used during initialization. bool None

layout_unclip_ratio Same as the parameter used during initialization. float|Tuple[float,float]|dict None

layout_merge_bboxes_mode Same as the parameter used during initialization. str|dict None

text_det_limit_side_len Same as the parameter used during initialization. int None

text_det_limit_type Same as the parameter used during initialization. str None

text_det_thresh Same as the parameter used during initialization. float None

text_det_box_thresh Same as the parameter used during initialization. float None

text_det_unclip_ratio Same as the parameter used during initialization. float None

text_rec_score_thresh Same as the parameter used during initialization. float None

seal_det_limit_side_len Same as the parameter used during initialization. int None

seal_det_limit_type Same as the parameter used during initialization. str None

seal_det_thresh Same as the parameter used during initialization. float None

seal_det_box_thresh Same as the parameter used during initialization. float None

seal_det_unclip_ratio Same as the parameter used during initialization. float None

seal_rec_score_thresh Same as the parameter used during initialization. float None

(3) Process the prediction results: each prediction result corresponds to a Result object, which supports printing, saving as image, or saving as a json file:

Method Description Parameter Type Parameter Description Default

print() Print result to terminal format_json bool Whether to format output as indented JSON True

indent int Indentation level to beautify the JSON output. Only effective when format_json=True 4

ensure_ascii bool Whether to escape non-ASCII characters to Unicode. When True, all non-ASCII characters are escaped. When False, original characters are retained. Only effective when format_json=True False

save_to_json() Save result as a JSON file save_path str Path to save the file. If a directory, the filename will be based on the input type None

indent int Indentation level for beautified JSON output. Only effective when format_json=True 4

ensure_ascii bool Whether to escape non-ASCII characters to Unicode. Only effective when format_json=True False

save_to_img() Save intermediate visualization results as PNG image files save_path str Path to save the file, supports directory or file path None

save_to_markdown() Save each page of an image or PDF file as a markdown file save_path str Path to save the file, supports directory or file path None

save_to_html() Save tables in the file as HTML format save_path str Path to save the file, supports directory or file path None

save_to_xlsx() Save tables in the file as XLSX format save_path str Path to save the file, supports directory or file path None

concatenate_markdown_pages() Concatenate multiple markdown pages into a single document markdown_list list List of markdown data for each page Returns the merged markdown text and image list

- Calling `print()` will print the result to the terminal. Explanation of the printed content: - `input_path`: `(str)` Input path of the image or PDF to be predicted - `page_index`: `(Union[int, None])` If input is a PDF, indicates the page number; otherwise `None` - `model_settings`: `(Dict[str, bool])` Model parameters configured for the pipeline - `use_doc_preprocessor`: `(bool)` Whether to enable document preprocessor sub-pipeline - `use_seal_recognition`: `(bool)` Whether to enable seal recognition sub-pipeline - `use_table_recognition`: `(bool)` Whether to enable table recognition sub-pipeline - `use_formula_recognition`: `(bool)` Whether to enable formula recognition sub-pipeline - `doc_preprocessor_res`: `(Dict[str, Union[List[float], str]])` Document preprocessing result dictionary, only exists if `use_doc_preprocessor=True` - `input_path`: `(str)` Image path accepted by document preprocessor, `None` if input is `numpy.ndarray` - `page_index`: `None` since input is `numpy.ndarray` - `model_settings`: `(Dict[str, bool])` Model configuration for the document preprocessor - `use_doc_orientation_classify`: `(bool)` Whether to enable document orientation classification - `use_doc_unwarping`: `(bool)` Whether to enable image unwarping - `angle`: `(int)` Predicted angle result if orientation classification is enabled - `parsing_res_list`: `(List[Dict])` List of parsed results, each item is a dictionary in reading order - `block_bbox`: `(np.ndarray)` Bounding box of the layout block - `block_label`: `(str)` Block label such as `text`, `table` - `block_content`: `(str)` Content within the layout block - `seg_start_flag`: `(bool)` Whether the block starts a paragraph - `seg_end_flag`: `(bool)` Whether the block ends a paragraph - `sub_label`: `(str)` Sub-label of the block, e.g., `title_text` - `sub_index`: `(int)` Sub-index of the block, used for markdown reconstruction - `index`: `(int)` Index of the block, used for layout sorting - `overall_ocr_res`: `(Dict[str, Union[List[str], List[float], numpy.ndarray]])` Dictionary of global OCR results - `input_path`: `(Union[str, None])` OCR sub-pipeline input path; `None` if input is `numpy.ndarray` - `page_index`: `None` since input is `numpy.ndarray` - `model_settings`: `(Dict)` OCR model configuration - `dt_polys`: `(List[numpy.ndarray])` List of polygons for text detection. Each box is a numpy array with shape (4, 2), dtype int16 - `dt_scores`: `(List[float])` Confidence scores for detection boxes - `text_det_params`: `(Dict[str, Dict[str, int, float]])` Text detection module parameters - `limit_side_len`: `(int)` Side length limit for image preprocessing - `limit_type`: `(str)` Limit processing method - `thresh`: `(float)` Threshold for text pixel classification - `box_thresh`: `(float)` Threshold for text detection boxes - `unclip_ratio`: `(float)` Unclip ratio for expanding boxes - `text_type`: `(str)` Text detection type, currently fixed as "general" - `text_type`: `(str)` Text detection type, currently fixed as "general" - `textline_orientation_angles`: `(List[int])` Orientation classification results for text lines - `text_rec_score_thresh`: `(float)` Threshold for text recognition filtering - `rec_texts`: `(List[str])` Recognized texts filtered by score threshold - `rec_scores`: `(List[float])` Recognition scores filtered by threshold - `rec_polys`: `(List[numpy.ndarray])` Filtered detection boxes, same format as `dt_polys` - `formula_res_list`: `(List[Dict[str, Union[numpy.ndarray, List[float], str]]])` List of formula recognition results - `rec_formula`: `(str)` Recognized formula string - `rec_polys`: `(numpy.ndarray)` Bounding box for the formula, shape (4, 2), dtype int16 - `formula_region_id`: `(int)` Region ID of the formula - `seal_res_list`: `(List[Dict[str, Union[numpy.ndarray, List[float], str]]])` List of seal recognition results - `input_path`: `(str)` Input path for the seal image - `page_index`: `None` since input is `numpy.ndarray` - `model_settings`: `(Dict)` Model configuration for seal recognition - `dt_polys`: `(List[numpy.ndarray])` Seal detection boxes, same format as `dt_polys` - `text_det_params`: `(Dict[str, Dict[str, int, float]])` Detection parameters, same as above - `text_type`: `(str)` Detection type, currently fixed as "seal" - `text_rec_score_thresh`: `(float)` Score threshold for recognition - `rec_texts`: `(List[str])` Recognized texts filtered by score - `rec_scores`: `(List[float])` Recognition scores filtered by threshold - `rec_polys`: `(List[numpy.ndarray])` Filtered seal boxes, same format as `dt_polys` - `rec_boxes`: `(numpy.ndarray)` Rectangle boxes, shape (n, 4), dtype int16 - `table_res_list`: `(List[Dict[str, Union[numpy.ndarray, List[float], str]]])` List of table recognition results - `cell_box_list`: `(List[numpy.ndarray])` Bounding boxes of table cells - `pred_html`: `(str)` Table as an HTML string - `table_ocr_pred`: `(dict)` OCR results for the table - `rec_polys`: `(List[numpy.ndarray])` Detected cell boxes - `rec_texts`: `(List[str])` Recognized texts for cells - `rec_scores`: `(List[float])` Confidence scores for cell recognition - `rec_boxes`: `(numpy.ndarray)` Rectangle boxes for detection, shape (n, 4), dtype int16 - Calling `save_to_json()` saves the above content to the specified `save_path`. If it’s a directory, the saved path will be `save_path/{your_img_basename}_res.json`. If it’s a file, it saves directly. Numpy arrays are converted to lists since JSON doesn't support them. - Calling `save_to_img()` saves visual results to the specified `save_path`. If a directory, various visualizations such as layout detection, OCR, and reading order are saved. If a file, only the last image is saved and others are overwritten. - Calling `save_to_markdown()` saves converted markdown files to `save_path/{your_img_basename}.md`. For PDF input, it's recommended to specify a directory to avoid file overwriting. - Calling `concatenate_markdown_pages()` merges multi-page markdown results from the `PP-StructureV3 pipeline` into a single document and returns the merged content. Additionally, you can access the prediction results and visual images through the following attributes:

Attribute Description

json Get the prediction result in json format

img Get visualized image results as a dict

markdown Get markdown results as a dict

- The `json` attribute returns the prediction result as a dictionary, which is consistent with the content saved using the `save_to_json()` method. - The `img` attribute returns the prediction result as a dictionary. The keys include `layout_det_res`, `overall_ocr_res`, `text_paragraphs_ocr_res`, `formula_res_region1`, `table_cell_img`, and `seal_res_region1`, each corresponding to a visualized `Image.Image` object for layout detection, OCR, text paragraph, formula, table, and seal results. If optional modules are not used, the dictionary only contains `layout_det_res`. - The `markdown` attribute returns the prediction result as a dictionary. The keys include `markdown_texts`, `markdown_images`, and `page_continuation_flags`, where the values represent the markdown text, displayed images (`Image.Image` objects), and a boolean tuple indicating whether the first and last elements of the current page are paragraph boundaries.

3. Development Integration / Deployment¶

If the pipeline meets your requirements for inference speed and accuracy, you can proceed with development integration or deployment.

If you want to directly use the pipeline in your Python project, refer to the example code in 2.2 Python script mode.

In addition, PaddleOCR provides two other deployment options described in detail below:

🚀 High-Performance Inference: In production environments, many applications have strict performance requirements (especially response speed) to ensure system efficiency and smooth user experience. PaddleOCR offers a high-performance inference option that deeply optimizes model inference and pre/post-processing for significant end-to-end acceleration. For detailed high-performance inference workflow, refer to High Performance Inference.

☁️ Service Deployment: Service-based deployment is common in production. It encapsulates the inference logic as a service, allowing clients to access it via network requests to obtain results. For detailed instructions on service deployment, refer to Service Deployment.

Below is the API reference and multi-language service invocation examples for basic service deployment:

API Reference

Main operations provided by the service:

HTTP method: POST

Request and response bodies are both JSON objects.

When the request is successful, the response status code is 200, and the response body contains:

Name Type Description

logId string UUID of the request

errorCode integer Error code, fixed to 0

errorMsg string Error message, fixed to "Success"

result object Operation result

When the request fails, the response body includes:

Name Type Description

logId string UUID of the request

errorCode integer Error code, same as HTTP status code

errorMsg string Error message

Main operation provided:

infer

Perform layout parsing.

POST /layout-parsing

Request body parameters:

Name Type Description Required

file string URL of image or PDF file accessible to the server, or base64-encoded file content. By default, only the first 10 pages of a PDF are processed.
To remove this limit, add the following to the pipeline config:
Serving: extra: max_num_input_imgs: null
Yes

fileType integer｜null File type. 0 for PDF, 1 for image. If omitted, the type is inferred from the URL. No

useDocOrientationClassify boolean | null Refer to the use_doc_orientation_classify parameter in the pipeline’s predict method. No

useDocUnwarping boolean | null Refer to the use_doc_unwarping parameter in the pipeline’s predict method. No

useTextlineOrientation boolean | null Refer to the use_textline_orientation parameter in the pipeline’s predict method. No

useSealRecognition boolean | null Refer to the use_seal_recognition parameter in the pipeline’s predict method. No

useTableRecognition boolean | null Refer to the use_table_recognition parameter in the pipeline’s predict method. No

useFormulaRecognition boolean | null Refer to the use_formula_recognition parameter in the pipeline’s predict method. No

layoutThreshold number | null Refer to the layout_threshold parameter in the pipeline’s predict method. No

layoutNms boolean | null Refer to the layout_nms parameter in the pipeline’s predict method. No

layoutUnclipRatio number | array | object | null Refer to the layout_unclip_ratio parameter in the pipeline’s predict method. No

layoutMergeBboxesMode string | object | null Refer to the layout_merge_bboxes_mode parameter in the pipeline’s predict method. No

textDetLimitSideLen integer | null Refer to the text_det_limit_side_len parameter in the pipeline’s predict method. No

textDetLimitType string | null Refer to the text_det_limit_type parameter in the pipeline’s predict method. No

textDetThresh number | null Refer to the text_det_thresh parameter in the pipeline’s predict method. No

textDetBoxThresh number | null Refer to the text_det_box_thresh parameter in the pipeline’s predict method. No

textDetUnclipRatio number | null Refer to the text_det_unclip_ratio parameter in the pipeline’s predict method. No

textRecScoreThresh number | null Refer to the text_rec_score_thresh parameter in the pipeline’s predict method. No

sealDetLimitSideLen integer | null Refer to the seal_det_limit_side_len parameter in the pipeline’s predict method. No

sealDetLimitType string | null Refer to the seal_det_limit_type parameter in the pipeline’s predict method. No

sealDetThresh number | null Refer to the seal_det_thresh parameter in the pipeline’s predict method. No

sealDetBoxThresh number | null Refer to the seal_det_box_thresh parameter in the pipeline’s predict method. No

sealDetUnclipRatio number | null Refer to the seal_det_unclip_ratio parameter in the pipeline’s predict method. No

sealRecScoreThresh number | null Refer to the seal_rec_score_thresh parameter in the pipeline’s predict method. No

useTableCellsOcrResults boolean Refer to the use_table_cells_ocr_results parameter in the pipeline’s predict method. No

useE2eWiredTableRecModel boolean Refer to the use_e2e_wired_table_rec_model parameter in the pipeline’s predict method. No

useE2eWirelessTableRecModel boolean Refer to the use_e2e_wireless_table_rec_model parameter in the pipeline’s predict method. No

When the request is successful, the result field of the response contains the following attributes:

Name Type Description

layoutParsingResults array Layout parsing results. The array length is 1 (for image input) or the number of processed pages (for PDF input). For PDF input, each element corresponds to one processed page.

dataInfo object Information about the input data.

Each element in layoutParsingResults is an object with the following attributes:

Name Type Description

prunedResult object A simplified version of the res field from the JSON output of the pipeline’s predict method, with input_path and page_index removed.

markdown object Markdown result.

outputImages object | null Refer to the pipeline’s img attribute. Images are JPEG encoded in Base64.

inputImage string | null Input image. JPEG encoded in Base64.

The markdown object has the following attributes:

Name Type Description

text string Markdown text.

images object Key-value pairs of image relative paths and base64-encoded image content.

isStart boolean Whether the first element on the current page is the start of a paragraph.

isEnd boolean Whether the last element on the current page is the end of a paragraph.

Multi-language Service Call Examples

Python

import base64 import requests import pathlib API_URL = "http://localhost:8080/layout-parsing" # Service URL image_path = "./demo.jpg" # Encode the local image to Base64 with open(image_path, "rb") as file: image_bytes = file.read() image_data = base64.b64encode(image_bytes).decode("ascii") payload = { "file": image_data, # Base64-encoded file content or file URL "fileType": 1, # File type, 1 indicates image file } # Call the API response = requests.post(API_URL, json=payload) # Handle the response data assert response.status_code == 200 result = response.json()["result"] for i, res in enumerate(result["layoutParsingResults"]): print(res["prunedResult"]) md_dir = pathlib.Path(f"markdown_{i}") md_dir.mkdir(exist_ok=True) (md_dir / "doc.md").write_text(res["markdown"]["text"]) for img_path, img in res["markdown"]["images"].items(): img_path = md_dir / img_path img_path.parent.mkdir(parents=True, exist_ok=True) img_path.write_bytes(base64.b64decode(img)) print(f"Markdown document saved at {md_dir / 'doc.md'}") for img_name, img in res["outputImages"].items(): img_path = f"{img_name}_{i}.jpg" with open(img_path, "wb") as f: f.write(base64.b64decode(img)) print(f"Output image saved at {img_path}")

4. Secondary Development¶

If the default model weights provided by the PP-StructureV3 pipeline do not meet your accuracy or speed requirements in your scenario, you can try fine-tuning the existing model using your own domain-specific or application-specific data to improve the performance of the PP-StructureV3 pipeline for your use case.

4.1 Model Fine-tuning¶

Since the PP-StructureV3 pipeline contains multiple modules, unsatisfactory results may originate from any individual module. You can analyze the problematic cases with poor extraction performance, visualize the images, identify the specific module causing the issue, and then refer to the fine-tuning tutorials linked in the table below to perform model fine-tuning.

Scenario Fine-tuning Module Fine-tuning Reference Link

Inaccurate layout detection, such as missing seals or tables Layout Detection Module Link

Inaccurate table structure recognition Table Structure Recognition Module Link

Inaccurate formula recognition Formula Recognition Module Link

Missing seal text detection Seal Text Detection Module Link

Missing text detection Text Detection Module Link

Incorrect text recognition results Text Recognition Module Link

Incorrect correction of vertical or rotated text lines Text Line Orientation Classification Module Link

Incorrect correction of full image orientation Document Image Orientation Classification Module Link

Inaccurate image distortion correction Text Image Correction Module Fine-tuning not supported yet

4.2 Model Deployment¶

Once you have completed fine-tuning with your private dataset, you will obtain the local model weights. You can then use these fine-tuned weights by customizing the pipeline configuration file.

Export the pipeline configuration file

You can call the export_paddlex_config_to_yaml method of the PPStructureV3 object in PaddleOCR to export the current pipeline configuration as a YAML file:

from paddleocr import PPStructureV3 pipeline = PPStructureV3() pipeline.export_paddlex_config_to_yaml("PP-StructureV3.yaml")

Modify the configuration file After obtaining the default pipeline configuration file, replace the corresponding path in the configuration with the local path of your fine-tuned model weights. For example:
...... SubModules: LayoutDetection: module_name: layout_detection model_name: PP-DocLayout_plus-L model_dir: null # Replace with the path to the fine-tuned layout detection model weights ...... SubPipelines: GeneralOCR: pipeline_name: OCR text_type: general use_doc_preprocessor: False use_textline_orientation: False SubModules: TextDetection: module_name: text_detection model_name: PP-OCRv5_server_det model_dir: null # Replace with the path to the fine-tuned text detection model weights limit_side_len: 960 limit_type: max max_side_limit: 4000 thresh: 0.3 box_thresh: 0.6 unclip_ratio: 1.5 TextRecognition: module_name: text_recognition model_name: PP-OCRv5_server_rec model_dir: null # Replace with the path to the fine-tuned text recognition model weights batch_size: 1 score_thresh: 0 ......

The pipeline configuration file not only includes parameters supported by the PaddleOCR CLI and Python API but also allows for more advanced configurations. For more details, refer to the corresponding pipeline usage tutorial in the PaddleX Pipeline Usage Overview, and adjust the configurations as needed based on your requirements.

Load the pipeline configuration file via CLI

After modifying the configuration file, specify the updated pipeline configuration path using the --paddlex_config parameter in the command line. PaddleOCR will load its content as the pipeline configuration. Example:

paddleocr pp_structurev3 --paddlex_config PP-StructureV3.yaml ...

Load the pipeline configuration file via Python API When initializing the pipeline object, you can pass the PaddleX pipeline configuration file path or a configuration dictionary using the paddlex_config parameter. PaddleOCR will load its content as the pipeline configuration. Example:

from paddleocr import PPStructureV3 pipeline = PPStructureV3(paddlex_config="PP-StructureV3.yaml")

May 20, 2025 May 20, 2025

Comments

Model	Download Link	mAP(0.5) (%)	GPU Inference Time (ms) [Standard Mode / High Performance Mode]	CPU Inference Time (ms) [Standard Mode / High Performance Mode]	Model Size (M)	Description
PP-DocLayout-L	Inference Model/Pretrained Model	90.4	34.6244 / 10.3945	510.57 / -	123.76 M	A high-precision layout area localization model trained on a self-built dataset containing Chinese and English papers, magazines, contracts, books, exams, and research reports using RT-DETR-L.
PP-DocLayout-M	Inference Model/Pretrained Model	75.2	13.3259 / 4.8685	44.0680 / 44.0680	22.578	A layout area localization model with balanced precision and efficiency, trained on a self-built dataset containing Chinese and English papers, magazines, contracts, books, exams, and research reports using PicoDet-L.
PP-DocLayout-S	Inference Model/Pretrained Model	70.9	8.3008 / 2.3794	10.0623 / 9.9296	4.834	A high-efficiency layout area localization model trained on a self-built dataset containing Chinese and English papers, magazines, contracts, books, exams, and research reports using PicoDet-S.

Model	Model Download Link	mAP(0.5) (%)	GPU Inference Time (ms) [Normal Mode / High-Performance Mode]	CPU Inference Time (ms) [Normal Mode / High-Performance Mode]	Model Storage Size (M)	Introduction
PicoDet-S_layout_3cls	Inference Model/Training Model	88.2	8.99 / 2.22	16.11 / 8.73	4.8	A high-efficiency layout area localization model trained on a self-built dataset of Chinese and English papers, magazines, and research reports using PicoDet-S.
PicoDet-L_layout_3cls	Inference Model/Training Model	89.0	13.05 / 4.50	41.30 / 41.30	22.6	A balanced efficiency and precision layout area localization model trained on a self-built dataset of Chinese and English papers, magazines, and research reports using PicoDet-L.

Model	Model Download Link	mAP(0.5) (%)	GPU Inference Time (ms) [Normal Mode / High-Performance Mode]	CPU Inference Time (ms) [Normal Mode / High-Performance Mode]	Model Storage Size (M)	Introduction
PicoDet-S_layout_17cls	Inference Model/Training Model	87.4	9.11 / 2.12	15.42 / 9.12	4.8	A high-efficiency layout area localization model trained on a self-built dataset of Chinese and English papers, magazines, and research reports using PicoDet-S.
PicoDet-L_layout_17cls	Inference Model/Training Model	89.0	13.50 / 4.69	43.32 / 43.32	22.6	A balanced efficiency and precision layout area localization model trained on a self-built dataset of Chinese and English papers, magazines, and research reports using PicoDet-L.
RT-DETR-H_layout_17cls	Inference Model/Training Model	98.3	115.29 / 104.09	995.27 / 995.27	470.2	A high-precision layout area localization model trained on a self-built dataset of Chinese and English papers, magazines, and research reports using RT-DETR-H.

Model	Download Link	mAP (%)	GPU Inference Time (ms) [Standard Mode / High Performance Mode]	CPU Inference Time (ms) [Standard Mode / High Performance Mode]	Model Size (M)	Description
RT-DETR-L_wired_table_cell_det	Inference Model/Pretrained Model	82.7	35.00 / 10.45	495.51 / 495.51	124M	RT-DETR is the first real-time end-to-end object detection model. Based on RT-DETR-L, the PaddlePaddle Vision Team pre-trained the model on a custom table cell detection dataset, achieving good performance for both wired and wireless tables.
RT-DETR-L_wireless_table_cell_det	Inference Model/Pretrained Model	82.7	35.00 / 10.45	495.51 / 495.51	124M

Model	Download Link	Detection Hmean (%)	GPU Inference Time (ms) [Standard Mode / High Performance Mode]	CPU Inference Time (ms) [Standard Mode / High Performance Mode]	Model Size (M)	Description
PP-OCRv4_server_det	Inference Model/Training Model	82.56	83.34 / 80.91	442.58 / 442.58	109	The server-side text detection model of PP-OCRv4, with higher accuracy, suitable for deployment on high-performance servers.
PP-OCRv4_mobile_det	Inference Model/Training Model	77.35	8.79 / 3.13	51.00 / 28.58	4.7	The mobile text detection model of PP-OCRv4, with higher efficiency, suitable for deployment on edge devices.
PP-OCRv3_mobile_det	Inference Model/Training Model	78.68	8.44 / 2.91	27.87 / 27.87	2.1	The mobile text detection model of PP-OCRv3, with higher efficiency, suitable for deployment on edge devices.
PP-OCRv3_server_det	Inference Model/Training Model	80.11	65.41 / 13.67	305.07 / 305.07	102.1	The server-side text detection model of PP-OCRv3, with higher accuracy, suitable for deployment on high-performance servers.

Model	Download Link	Chinese Avg Accuracy (%)	English Avg Accuracy (%)	Traditional Chinese Avg Accuracy (%)	Japanese Avg Accuracy (%)	GPU Inference Time (ms) [Standard Mode / High Performance Mode]	CPU Inference Time (ms) [Standard Mode / High Performance Mode]	Model Size (M)	Description
PP-OCRv5_server_rec	Inference Model/Pretrained Model	86.38	64.70	93.29	60.35	-	-	205 M	PP-OCRv5_server_rec is a new-generation text recognition model. It efficiently and accurately supports four major languages: Simplified Chinese, Traditional Chinese, English, and Japanese, as well as handwriting, vertical text, pinyin, and rare characters, offering robust and efficient support for document understanding.
PP-OCRv5_mobile_rec	Inference Model/Pretrained Model	81.29	66.00	83.55	54.65	-	-	136 M	PP-OCRv5_mobile_rec is a new-generation text recognition model. It efficiently and accurately supports four major languages: Simplified Chinese, Traditional Chinese, English, and Japanese, as well as handwriting, vertical text, pinyin, and rare characters, offering robust and efficient support for document understanding.

Model	Download Link	Avg Accuracy (%)	GPU Inference Time (ms) [Standard Mode / High Performance Mode]	CPU Inference Time (ms) [Standard Mode / High Performance Mode]	Model Size (M)	Description
PP-OCRv4_server_rec_doc	Inference Model/Pretrained Model	86.58	6.65 / 2.38	32.92 / 32.92	91 M	Based on PP-OCRv4_server_rec, trained on additional Chinese documents and PP-OCR mixed data. It supports over 15,000 characters including Traditional Chinese, Japanese, and special symbols, enhancing both document-specific and general text recognition accuracy.
PP-OCRv4_mobile_rec	Inference Model/Pretrained Model	83.28	4.82 / 1.20	16.74 / 4.64	11 M	Lightweight model of PP-OCRv4 with high inference efficiency, suitable for deployment on various edge devices.
PP-OCRv4_server_rec	Inference Model/Pretrained Model	85.19	6.58 / 2.43	33.17 / 33.17	87 M	Server-side model of PP-OCRv4 with high recognition accuracy, suitable for deployment on various servers.
PP-OCRv3_mobile_rec	Inference Model/Pretrained Model	75.43	5.87 / 1.19	9.07 / 4.28	11 M	Lightweight model of PP-OCRv3 with high inference efficiency, suitable for deployment on various edge devices.

Model	Download Link	Avg Accuracy (%)	GPU Inference Time (ms) [Standard Mode / High Performance Mode]	CPU Inference Time (ms) [Standard Mode / High Performance Mode]	Model Size (M)	Description
en_PP-OCRv4_mobile_rec	Inference Model/Pretrained Model	70.39	4.81 / 0.75	16.10 / 5.31	6.8 M	Ultra-lightweight English recognition model trained on PP-OCRv4, supporting English and number recognition.
en_PP-OCRv3_mobile_rec	Inference Model/Pretrained Model	70.69	5.44 / 0.75	8.65 / 5.57	7.8 M	Ultra-lightweight English recognition model trained on PP-OCRv3, supporting English and number recognition.

Model	Model Download Link	Recognition Avg Accuracy(%)	GPU Inference Time (ms) [Normal / High Performance]	CPU Inference Time (ms) [Normal / High Performance]	Model Size (M)	Description
korean_PP-OCRv3_mobile_rec	Inference Model/Pretrained Model	60.21	5.40 / 0.97	9.11 / 4.05	8.6 M	An ultra-lightweight Korean text recognition model trained based on PP-OCRv3, supporting Korean and digits recognition
japan_PP-OCRv3_mobile_rec	Inference Model/Pretrained Model	45.69	5.70 / 1.02	8.48 / 4.07	8.8 M	An ultra-lightweight Japanese text recognition model trained based on PP-OCRv3, supporting Japanese and digits recognition
chinese_cht_PP-OCRv3_mobile_rec	Inference Model/Pretrained Model	82.06	5.90 / 1.28	9.28 / 4.34	9.7 M	An ultra-lightweight Traditional Chinese text recognition model trained based on PP-OCRv3, supporting Traditional Chinese and digits recognition
te_PP-OCRv3_mobile_rec	Inference Model/Pretrained Model	95.88	5.42 / 0.82	8.10 / 6.91	7.8 M	An ultra-lightweight Telugu text recognition model trained based on PP-OCRv3, supporting Telugu and digits recognition
ka_PP-OCRv3_mobile_rec	Inference Model/Pretrained Model	96.96	5.25 / 0.79	9.09 / 3.86	8.0 M	An ultra-lightweight Kannada text recognition model trained based on PP-OCRv3, supporting Kannada and digits recognition
ta_PP-OCRv3_mobile_rec	Inference Model/Pretrained Model	76.83	5.23 / 0.75	10.13 / 4.30	8.0 M	An ultra-lightweight Tamil text recognition model trained based on PP-OCRv3, supporting Tamil and digits recognition
latin_PP-OCRv3_mobile_rec	Inference Model/Pretrained Model	76.93	5.20 / 0.79	8.83 / 7.15	7.8 M	An ultra-lightweight Latin text recognition model trained based on PP-OCRv3, supporting Latin and digits recognition
arabic_PP-OCRv3_mobile_rec	Inference Model/Pretrained Model	73.55	5.35 / 0.79	8.80 / 4.56	7.8 M	An ultra-lightweight Arabic script recognition model trained based on PP-OCRv3, supporting Arabic script and digits recognition
cyrillic_PP-OCRv3_mobile_rec	Inference Model/Pretrained Model	94.28	5.23 / 0.76	8.89 / 3.88	7.9 M	An ultra-lightweight Cyrillic script recognition model trained based on PP-OCRv3, supporting Cyrillic script and digits recognition
devanagari_PP-OCRv3_mobile_rec	Inference Model/Pretrained Model	96.44	5.22 / 0.79	8.56 / 4.06	7.9 M	An ultra-lightweight Devanagari script recognition model trained based on PP-OCRv3, supporting Devanagari script and digits recognition

Model	Model Download Link	En-BLEU(%)	Zh-BLEU(%)	GPU Inference Time (ms) [Normal Mode / High-Performance Mode]	CPU Inference Time (ms) [Normal Mode / High-Performance Mode]	Model Storage Size (M)	Introduction
UniMERNet	Inference Model/Training Model	85.91	43.50	2266.96/-	-/-	1.53 G	UniMERNet is a formula recognition model developed by Shanghai AI Lab. It uses Donut Swin as the encoder and MBartDecoder as the decoder. The model is trained on a dataset of one million samples, including simple formulas, complex formulas, scanned formulas, and handwritten formulas, significantly improving the recognition accuracy of real-world formulas.
PP-FormulaNet-S	Inference Model/Training Model	87.00	45.71	202.25/-	-/-	224 M	PP-FormulaNet is an advanced formula recognition model developed by the Baidu PaddlePaddle Vision Team. The PP-FormulaNet-S version uses PP-HGNetV2-B4 as its backbone network. Through parallel masking and model distillation techniques, it significantly improves inference speed while maintaining high recognition accuracy, making it suitable for applications requiring fast inference. The PP-FormulaNet-L version, on the other hand, uses Vary_VIT_B as its backbone network and is trained on a large-scale formula dataset, showing significant improvements in recognizing complex formulas compared to PP-FormulaNet-S.
PP-FormulaNet-L	Inference Model/Training Model	90.36	45.78	1976.52/-	-/-	695 M
PP-FormulaNet_plus-S	Inference Model/Training Model	88.71	53.32	191.69/-	-/-	248 M	PP-FormulaNet_plus is an enhanced version of the formula recognition model developed by the Baidu PaddlePaddle Vision Team, building upon the original PP-FormulaNet. Compared to the original version, PP-FormulaNet_plus utilizes a more diverse formula dataset during training, including sources such as Chinese dissertations, professional books, textbooks, exam papers, and mathematics journals. This expansion significantly improves the model’s recognition capabilities. Among the models, PP-FormulaNet_plus-M and PP-FormulaNet_plus-L have added support for Chinese formulas and increased the maximum number of predicted tokens for formulas from 1,024 to 2,560, greatly enhancing the recognition performance for complex formulas. Meanwhile, the PP-FormulaNet_plus-S model focuses on improving the recognition of English formulas. With these improvements, the PP-FormulaNet_plus series models perform exceptionally well in handling complex and diverse formula recognition tasks.
PP-FormulaNet_plus-M	Inference Model/Training Model	91.45	89.76	1301.56/-	-/-	592 M
PP-FormulaNet_plus-L	Inference Model/Training Model	92.22	90.64	1745.25/-	-/-	698 M
LaTeX_OCR_rec	Inference Model/Training Model	74.55	39.96	1244.61/-	-/-	99 M	LaTeX-OCR is a formula recognition algorithm based on an autoregressive large model. It uses Hybrid ViT as the backbone network and a transformer as the decoder, significantly improving the accuracy of formula recognition.

Model	Model Download Link	Detection Hmean (%)	GPU Inference Time (ms) [Normal / High Performance]	CPU Inference Time (ms) [Normal / High Performance]	Model Size (M)	Description
PP-OCRv4_server_seal_det	Inference Model/Pretrained Model	98.21	74.75 / 67.72	382.55 / 382.55	109	Server-side seal text detection model based on PP-OCRv4, offering higher accuracy and suitable for deployment on high-performance servers
PP-OCRv4_mobile_seal_det	Inference Model/Pretrained Model	96.47	7.82 / 3.09	48.28 / 23.97	4.6	Mobile-side seal text detection model based on PP-OCRv4, offering higher efficiency and suitable for edge-side deployment

Mode	GPU Configuration	CPU Configuration	Acceleration Technology Combination
Normal Mode	FP32 Precision / No TRT Acceleration	FP32 Precision / 8 Threads	PaddleInference
High-Performance Mode	Optimal combination of pre-selected precision types and acceleration strategies	FP32 Precision / 8 Threads	Pre-selected optimal backend (Paddle/OpenVINO/TRT, etc.)

Parameter	Description	Type	Default
`input`	Data to be predicted. Required. Supports multiple input types. Python Var: e.g., `numpy.ndarray` representing image data str: e.g., local path to image or PDF file: `/root/data/img.jpg`; URL, e.g., online image or PDF: example; local directory: directory containing images to predict, e.g., `/root/data/` (currently, directories with PDFs are not supported; PDFs must be specified by file path) List: list elements must be one of the above types, e.g., `[numpy.ndarray, numpy.ndarray]`, `["/root/data/img1.jpg", "/root/data/img2.jpg"]`, `["/root/data1", "/root/data2"]`	`Python Var\|str\|list`
`save_path`	Path to save inference results. If set to `None`, results will not be saved locally.	`str`	`None`
`layout_detection_model_name`	Name of the layout detection model. If set to `None`, the default model will be used.	`str`	`None`
`layout_detection_model_dir`	Directory path of the layout detection model. If set to `None`, the official model will be downloaded.	`str`	`None`
`layout_threshold`	Score threshold for the layout model. float: any value between `0-1` dict: `{0:0.1}`, where key is class ID and value is the threshold for that class None: if set to `None`, the default value is used, which is `0.5`	`float\|dict`	`None`
`layout_nms`	Whether to apply NMS post-processing for layout detection model.	`bool`	`None`
`layout_unclip_ratio`	Unclip ratio for detected boxes in layout detection model. float: any float > `0` Tuple[float,float]: separate ratios for width and height dict: key is int (class ID), value is tuple, e.g., `{0: (1.1, 2.0)}` means class 0 boxes will be expanded 1.1x in width, 2.0x in height None: if set to `None`, default is `1.0`	`float\|Tuple[float,float]\|dict`	`None`
`layout_merge_bboxes_mode`	Merge mode for overlapping boxes in layout detection. str: `large`, `small`, `union`, for keeping larger box, smaller box, or both dict: key is int (class ID), value is str, e.g., `{0: "large", 2: "small"}` None: if set to `None`, default is `large`	`str\|dict`	`None`
`chart_recognition_model_name`	Name of the chart recognition model. If set to `None`, the default model will be used.	`str`	`None`
`chart_recognition_model_dir`	Directory path of the chart recognition model. If set to `None`, the official model will be downloaded.	`str`	`None`
`chart_recognition_batch_size`	Batch size for the chart recognition model. If set to `None`, the default batch size is `1`.	`int`	`None`
`region_detection_model_name`	Name of the region detection model. If set to `None`, the default model will be used.	`str`	`None`
`region_detection_model_dir`	Directory path of the region detection model. If set to `None`, the official model will be downloaded.	`str`	`None`
`doc_orientation_classify_model_name`	Name of the document orientation classification model. If set to `None`, the default model will be used.	`str`	`None`
`doc_orientation_classify_model_dir`	Directory path of the document orientation classification model. If set to `None`, the official model will be downloaded.	`str`	`None`
`doc_unwarping_model_name`	Name of the document unwarping model. If set to `None`, the default model will be used.	`str`	`None`
`doc_unwarping_model_dir`	Directory path of the document unwarping model. If set to `None`, the official model will be downloaded.	`str`	`None`
`text_detection_model_name`	Name of the text detection model. If set to `None`, the default model will be used.	`str`	`None`
`text_detection_model_dir`	Directory path of the text detection model. If set to `None`, the official model will be downloaded.	`str`	`None`
`text_det_limit_side_len`	Maximum side length limit for text detection. int: any integer > `0`; None: if set to `None`, the default value will be `960`;	`int`	`None`
`text_det_limit_type`	str: supports `min` and `max`; `min` means ensuring the shortest side of the image is not less than `det_limit_side_len`, `max` means the longest side does not exceed `limit_side_len` None: if set to `None`, the default value will be `max`.	`str`	`None`
`text_det_thresh`	Pixel threshold for detection. Pixels with scores above this value in the probability map are considered text. float: any float > `0` None: if set to `None`, default is `0.3`	`float`	`None`
`text_det_box_thresh`	Box threshold. A bounding box is considered text if the average score of pixels inside is greater than this value. float: any float > `0` None: if set to `None`, default is `0.6`	`float`	`None`
`text_det_unclip_ratio`	Expansion ratio for text detection. The higher the value, the larger the expansion area. float: any float > `0` None: if set to `None`, default is `2.0`	`float`	`None`
`textline_orientation_model_name`	Name of the text line orientation model. If set to `None`, the default model will be used.	`str`	`None`
`textline_orientation_model_dir`	Directory of the text line orientation model. If set to `None`, the official model will be downloaded.	`str`	`None`
`textline_orientation_batch_size`	Batch size for the text line orientation model. If set to `None`, default is `1`.	`int`	`None`
`text_recognition_model_name`	Name of the text recognition model. If set to `None`, the default model will be used.	`str`	`None`
`text_recognition_model_dir`	Directory of the text recognition model. If set to `None`, the official model will be downloaded.	`str`	`None`
`text_recognition_batch_size`	Batch size for text recognition. If set to `None`, default is `1`.	`int`	`None`
`text_rec_score_thresh`	Score threshold for text recognition. Only results above this value will be kept. float: any float > `0` None: if set to `None`, default is `0.0` (no threshold)	`float`	`None`
`table_classification_model_name`	Name of the table classification model. If set to `None`, the default model will be used.	`str`	`None`
`table_classification_model_dir`	Directory of the table classification model. If set to `None`, the official model will be downloaded.	`str`	`None`
`wired_table_structure_recognition_model_name`	Name of the wired table structure recognition model. If set to `None`, the default model will be used.	`str`	`None`
`wired_table_structure_recognition_model_dir`	Directory of the wired table structure recognition model. If set to `None`, the official model will be downloaded.	`str`	`None`
`wireless_table_structure_recognition_model_name`	Name of the wireless table structure recognition model. If set to `None`, the default model will be used.	`str`	`None`
`wireless_table_structure_recognition_model_dir`	Directory of the wireless table structure recognition model. If set to `None`, the official model will be downloaded.	`str`	`None`
`wired_table_cells_detection_model_name`	Name of the wired table cell detection model. If set to `None`, the default model will be used.	`str`	`None`
`wired_table_cells_detection_model_dir`	Directory of the wired table cell detection model. If set to `None`, the official model will be downloaded.	`str`	`None`
`wireless_table_cells_detection_model_name`	Name of the wireless table cell detection model. If set to `None`, the default model will be used.	`str`	`None`
`wireless_table_cells_detection_model_dir`	Directory of the wireless table cell detection model. If set to `None`, the official model will be downloaded.	`str`	`None`
`seal_text_detection_model_name`	Name of the seal text detection model. If set to `None`, the default model will be used.	`str`	`None`
`seal_text_detection_model_dir`	Directory of the seal text detection model. If set to `None`, the official model will be downloaded.	`str`	`None`
`seal_det_limit_side_len`	Image side length limit for seal text detection. int: any integer > `0`; None: if set to `None`, the default is `736`;	`int`	`None`
`seal_det_limit_type`	Limit type for image side in seal text detection. str: supports `min` and `max`; `min` ensures shortest side ≥ `det_limit_side_len`, `max` ensures longest side ≤ `limit_side_len` None: if set to `None`, default is `min`;	`str`	`None`
`seal_det_thresh`	Pixel threshold. Pixels with scores above this value in the probability map are considered text. float: any float > `0` None: if set to `None`, default is `0.2`	`float`	`None`
`seal_det_box_thresh`	Box threshold. Boxes with average pixel scores above this value are considered text regions. float: any float > `0` None: if set to `None`, default is `0.6`	`float`	`None`
`seal_det_unclip_ratio`	Expansion ratio for seal text detection. Higher value means larger expansion area. float: any float > `0` None: if set to `None`, default is `0.5`	`float`	`None`
`seal_text_recognition_model_name`	Name of the seal text recognition model. If set to `None`, the default model will be used.	`str`	`None`
`seal_text_recognition_model_dir`	Directory of the seal text recognition model. If set to `None`, the official model will be downloaded.	`str`	`None`
`seal_text_recognition_batch_size`	Batch size for seal text recognition. If set to `None`, default is `1`.	`int`	`None`
`seal_rec_score_thresh`	Recognition score threshold. Text results above this value will be kept. float: any float > `0` None: if set to `None`, default is `0.0` (no threshold)	`float`	`None`
`formula_recognition_model_name`	Name of the formula recognition model. If set to `None`, the default model will be used.	`str`	`None`
`formula_recognition_model_dir`	Directory of the formula recognition model. If set to `None`, the official model will be downloaded.	`str`	`None`
`formula_recognition_batch_size`	Batch size of the formula recognition model. If set to `None`, the default is `1`.	`int`	`None`
`use_doc_orientation_classify`	Whether to enable document orientation classification. If set to `None`, default is `True`.	`bool`	`None`
`use_doc_unwarping`	Whether to enable document unwarping. If set to `None`, default is `True`.	`bool`	`None`
`use_seal_recognition`	Whether to enable seal recognition subpipeline. If set to `None`, default is `True`.	`bool`	`None`
`use_table_recognition`	Whether to enable table recognition subpipeline. If set to `None`, default is `True`.	`bool`	`None`
`use_formula_recognition`	Whether to enable formula recognition subpipeline. If set to `None`, default is `True`.	`bool`	`None`
`use_chart_recognition`	Whether to enable chart recognition model. If set to `None`, default is `True`.	`bool`	`None`
`use_region_detection`	Whether to enable region detection submodule for document images. If set to `None`, default is `True`.	`bool`	`None`
`device`	Device for inference. You can specify a device ID. CPU: e.g., `cpu` GPU: e.g., `gpu:0` means GPU 0 NPU: e.g., `npu:0` means NPU 0 XPU: e.g., `xpu:0` means XPU 0 MLU: e.g., `mlu:0` means MLU 0 DCU: e.g., `dcu:0` means DCU 0 None: If set to `None`, GPU 0 will be used by default if available; otherwise, CPU will be used.	`str`	`None`
`enable_hpi`	Whether to enable high performance inference.	`bool`	`False`
`use_tensorrt`	Whether to use TensorRT for inference acceleration.	`bool`	`False`
`min_subgraph_size`	Minimum subgraph size for optimizing subgraph execution.	`int`	`3`
`precision`	Computation precision, e.g., fp32, fp16.	`str`	`fp32`
`enable_mkldnn`	Whether to enable MKL-DNN. If set to `None`, enabled by default.	`bool`	`None`
`cpu_threads`	Number of threads to use when inferring on CPU.	`int`	`8`
`paddlex_config`	Path to the PaddleX pipeline configuration file.	`str`	`None`

Parameter	Description	Type	Default
`input`	Input data to be predicted. Required. Supports multiple types: Python Var: Image data represented as `numpy.ndarray` str: Local path to image or PDF file, e.g., `/root/data/img.jpg`; URL to image or PDF, e.g., example; directory containing image files, e.g., `/root/data/` (directories with PDFs are not supported, use full file path for PDFs) List: Elements can be any of the above types, e.g., `[numpy.ndarray, numpy.ndarray]`, `["/root/data/img1.jpg", "/root/data/img2.jpg"]`, `["/root/data1", "/root/data2"]`	`Python Var\|str\|list`
`device`	Same as the parameter used during initialization.	`str`	`None`
`use_doc_orientation_classify`	Whether to use document orientation classification during inference.	`bool`	`None`
`use_doc_unwarping`	Whether to use document image unwarping during inference.	`bool`	`None`
`use_textline_orientation`	Whether to use textline orientation classification during inference.	`bool`	`None`
`use_seal_recognition`	Whether to use the seal recognition sub-pipeline during inference.	`bool`	`None`
`use_table_recognition`	Whether to use the table recognition sub-pipeline during inference.	`bool`	`None`
`use_formula_recognition`	Whether to use the formula recognition sub-pipeline during inference.	`bool`	`None`
`layout_threshold`	Same as the parameter used during initialization.	`float\|dict`	`None`
`layout_nms`	Same as the parameter used during initialization.	`bool`	`None`
`layout_unclip_ratio`	Same as the parameter used during initialization.	`float\|Tuple[float,float]\|dict`	`None`
`layout_merge_bboxes_mode`	Same as the parameter used during initialization.	`str\|dict`	`None`
`text_det_limit_side_len`	Same as the parameter used during initialization.	`int`	`None`
`text_det_limit_type`	Same as the parameter used during initialization.	`str`	`None`
`text_det_thresh`	Same as the parameter used during initialization.	`float`	`None`
`text_det_box_thresh`	Same as the parameter used during initialization.	`float`	`None`
`text_det_unclip_ratio`	Same as the parameter used during initialization.	`float`	`None`
`text_rec_score_thresh`	Same as the parameter used during initialization.	`float`	`None`
`seal_det_limit_side_len`	Same as the parameter used during initialization.	`int`	`None`
`seal_det_limit_type`	Same as the parameter used during initialization.	`str`	`None`
`seal_det_thresh`	Same as the parameter used during initialization.	`float`	`None`
`seal_det_box_thresh`	Same as the parameter used during initialization.	`float`	`None`
`seal_det_unclip_ratio`	Same as the parameter used during initialization.	`float`	`None`
`seal_rec_score_thresh`	Same as the parameter used during initialization.	`float`	`None`

Method	Description	Parameter	Type	Parameter Description	Default
`print()`	Print result to terminal	`format_json`	`bool`	Whether to format output as indented `JSON`	`True`
		`indent`	`int`	Indentation level to beautify the `JSON` output. Only effective when `format_json=True`	4
		`ensure_ascii`	`bool`	Whether to escape non-`ASCII` characters to `Unicode`. When `True`, all non-ASCII characters are escaped. When `False`, original characters are retained. Only effective when `format_json=True`	`False`
`save_to_json()`	Save result as a JSON file	`save_path`	`str`	Path to save the file. If a directory, the filename will be based on the input type	None
		`indent`	`int`	Indentation level for beautified `JSON` output. Only effective when `format_json=True`	4
		`ensure_ascii`	`bool`	Whether to escape non-`ASCII` characters to `Unicode`. Only effective when `format_json=True`	`False`
`save_to_img()`	Save intermediate visualization results as PNG image files	`save_path`	`str`	Path to save the file, supports directory or file path	None
`save_to_markdown()`	Save each page of an image or PDF file as a markdown file	`save_path`	`str`	Path to save the file, supports directory or file path	None
`save_to_html()`	Save tables in the file as HTML format	`save_path`	`str`	Path to save the file, supports directory or file path	None
`save_to_xlsx()`	Save tables in the file as XLSX format	`save_path`	`str`	Path to save the file, supports directory or file path	None
`concatenate_markdown_pages()`	Concatenate multiple markdown pages into a single document	`markdown_list`	`list`	List of markdown data for each page	Returns the merged markdown text and image list

Attribute	Description
`json`	Get the prediction result in `json` format
`img`	Get visualized image results as a `dict`
`img`	Get visualized image results as a `dict`
`markdown`	Get markdown results as a `dict`

Name	Type	Description
`logId`	`string`	UUID of the request
`errorCode`	`integer`	Error code, fixed to `0`
`errorMsg`	`string`	Error message, fixed to `"Success"`
`result`	`object`	Operation result

Name	Type	Description	Required
`file`	`string`	URL of image or PDF file accessible to the server, or base64-encoded file content. By default, only the first 10 pages of a PDF are processed. To remove this limit, add the following to the pipeline config: `Serving: extra: max_num_input_imgs: null`	Yes
`fileType`	`integer`｜`null`	File type. `0` for PDF, `1` for image. If omitted, the type is inferred from the URL.	No
`useDocOrientationClassify`	`boolean` \| `null`	Refer to the `use_doc_orientation_classify` parameter in the pipeline’s `predict` method.	No
`useDocUnwarping`	`boolean` \| `null`	Refer to the `use_doc_unwarping` parameter in the pipeline’s `predict` method.	No
`useTextlineOrientation`	`boolean` \| `null`	Refer to the `use_textline_orientation` parameter in the pipeline’s `predict` method.	No
`useSealRecognition`	`boolean` \| `null`	Refer to the `use_seal_recognition` parameter in the pipeline’s `predict` method.	No
`useTableRecognition`	`boolean` \| `null`	Refer to the `use_table_recognition` parameter in the pipeline’s `predict` method.	No
`useFormulaRecognition`	`boolean` \| `null`	Refer to the `use_formula_recognition` parameter in the pipeline’s `predict` method.	No
`layoutThreshold`	`number` \| `null`	Refer to the `layout_threshold` parameter in the pipeline’s `predict` method.	No
`layoutNms`	`boolean` \| `null`	Refer to the `layout_nms` parameter in the pipeline’s `predict` method.	No
`layoutUnclipRatio`	`number` \| `array` \| `object` \| `null`	Refer to the `layout_unclip_ratio` parameter in the pipeline’s `predict` method.	No
`layoutMergeBboxesMode`	`string` \| `object` \| `null`	Refer to the `layout_merge_bboxes_mode` parameter in the pipeline’s `predict` method.	No
`textDetLimitSideLen`	`integer` \| `null`	Refer to the `text_det_limit_side_len` parameter in the pipeline’s `predict` method.	No
`textDetLimitType`	`string` \| `null`	Refer to the `text_det_limit_type` parameter in the pipeline’s `predict` method.	No
`textDetThresh`	`number` \| `null`	Refer to the `text_det_thresh` parameter in the pipeline’s `predict` method.	No
`textDetBoxThresh`	`number` \| `null`	Refer to the `text_det_box_thresh` parameter in the pipeline’s `predict` method.	No
`textDetUnclipRatio`	`number` \| `null`	Refer to the `text_det_unclip_ratio` parameter in the pipeline’s `predict` method.	No
`textRecScoreThresh`	`number` \| `null`	Refer to the `text_rec_score_thresh` parameter in the pipeline’s `predict` method.	No
`sealDetLimitSideLen`	`integer` \| `null`	Refer to the `seal_det_limit_side_len` parameter in the pipeline’s `predict` method.	No
`sealDetLimitType`	`string` \| `null`	Refer to the `seal_det_limit_type` parameter in the pipeline’s `predict` method.	No
`sealDetThresh`	`number` \| `null`	Refer to the `seal_det_thresh` parameter in the pipeline’s `predict` method.	No
`sealDetBoxThresh`	`number` \| `null`	Refer to the `seal_det_box_thresh` parameter in the pipeline’s `predict` method.	No
`sealDetUnclipRatio`	`number` \| `null`	Refer to the `seal_det_unclip_ratio` parameter in the pipeline’s `predict` method.	No
`sealRecScoreThresh`	`number` \| `null`	Refer to the `seal_rec_score_thresh` parameter in the pipeline’s `predict` method.	No
`useTableCellsOcrResults`	`boolean`	Refer to the `use_table_cells_ocr_results` parameter in the pipeline’s `predict` method.	No
`useE2eWiredTableRecModel`	`boolean`	Refer to the `use_e2e_wired_table_rec_model` parameter in the pipeline’s `predict` method.	No
`useE2eWirelessTableRecModel`	`boolean`	Refer to the `use_e2e_wireless_table_rec_model` parameter in the pipeline’s `predict` method.	No

Name	Type	Description
`layoutParsingResults`	`array`	Layout parsing results. The array length is 1 (for image input) or the number of processed pages (for PDF input). For PDF input, each element corresponds to one processed page.
`dataInfo`	`object`	Information about the input data.

Name	Type	Description
`prunedResult`	`object`	A simplified version of the `res` field from the JSON output of the pipeline’s `predict` method, with `input_path` and `page_index` removed.
`markdown`	`object`	Markdown result.
`outputImages`	`object` \| `null`	Refer to the pipeline’s `img` attribute. Images are JPEG encoded in Base64.
`inputImage`	`string` \| `null`	Input image. JPEG encoded in Base64.

Name	Type	Description
`text`	`string`	Markdown text.
`images`	`object`	Key-value pairs of image relative paths and base64-encoded image content.
`isStart`	`boolean`	Whether the first element on the current page is the start of a paragraph.
`isEnd`	`boolean`	Whether the last element on the current page is the end of a paragraph.

Scenario	Fine-tuning Module	Fine-tuning Reference Link
Inaccurate layout detection, such as missing seals or tables	Layout Detection Module	Link
Inaccurate table structure recognition	Table Structure Recognition Module	Link
Inaccurate formula recognition	Formula Recognition Module	Link
Missing seal text detection	Seal Text Detection Module	Link
Missing text detection	Text Detection Module	Link
Incorrect text recognition results	Text Recognition Module	Link
Incorrect correction of vertical or rotated text lines	Text Line Orientation Classification Module	Link
Incorrect correction of full image orientation	Document Image Orientation Classification Module	Link
Inaccurate image distortion correction	Text Image Correction Module	Fine-tuning not supported yet