Seal Text Recognition Pipeline Tutorial

1. Introduction to Seal Text Recognition Pipeline

Seal text recognition is a technology that automatically extracts and recognizes the content of seals from documents or images. The recognition of seal text is part of document processing and has many applications in various scenarios, such as contract comparison, warehouse entry and exit review, and invoice reimbursement review.

The seal text recognition pipeline is used to recognize the text content of seals, extracting the text information from seal images and outputting it in text form. This pipeline integrates the industry-renowned end-to-end OCR system PP-OCRv4, supporting the detection and recognition of curved seal text. Additionally, this pipeline integrates an optional layout region localization module, which can accurately locate the layout position of the seal within the entire document. It also includes optional document image orientation correction and distortion correction functions. Based on this pipeline, millisecond-level accurate text content prediction can be achieved on a CPU. This pipeline also provides flexible service deployment methods, supporting the use of multiple programming languages on various hardware. Moreover, it offers custom development capabilities, allowing you to train and fine-tune on your own dataset based on this pipeline, and the trained model can be seamlessly integrated.

The seal text recognition pipeline includes a seal text detection module and a text recognition module, as well as optional layout detection module, document image orientation classification module, and text image correction module.

• Seal Text Detection Module
• Text Recognition Module
• Layout Detection Module (Optional)
• Document Image Orientation Classification Module (Optional)
• Text Image Unwarping Module (Optional)

If you prioritize model accuracy, choose a model with higher accuracy. If you prioritize inference speed, choose a model with faster inference speed. If you prioritize model storage size, choose a model with smaller storage size.

Layout Region Detection Module (Optional):

• 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

• Layout detection model, including 23 common categories: document title, paragraph title, text, page number, abstract, table of contents, references, footnotes, header, footer, algorithm, formula, formula number, image, chart title, table, table title, seal, chart title, chart, header image, footer image, sidebar text

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-L	Inference Model/Training 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/Training 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/Training 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 13 full models. It includes multiple predefined models of different categories, among which there are 10 models specifically for the seal category. Apart from the three core models mentioned above, the remaining models are listed as follows:

👉 Details of Model List

* 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.
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.

* 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.

Document Image Orientation Classification Module (Optional):

Model	Model Download Link	Top-1 Acc (%)	GPU Inference Time (ms) [Normal Mode / High-Performance Mode]	CPU Inference Time (ms) [Normal Mode / High-Performance Mode]	Model Storage Size (M)	Description
PP-LCNet_x1_0_doc_ori	Inference Model/Training Model	99.06	2.31 / 0.43	3.37 / 1.27	7	A document image classification model based on PP-LCNet_x1_0, containing four categories: 0 degrees, 90 degrees, 180 degrees, and 270 degrees

Text Image Correction Module (Optional):

Model	Model Download Link	CER	Model Storage Size (M)	Description
UVDoc	Inference Model/Training Model	0.179	30.3 M	High-precision text image correction model

Text Detection Module:

Model	Model Download Link	Detection Hmean (%)	GPU Inference Time (ms) [Normal Mode / High-Performance Mode]	CPU Inference Time (ms) [Normal Mode / High-Performance Mode]	Model Storage Size (M)	Description
PP-OCRv4_server_seal_det	Inference Model/Training Model	98.21	74.75 / 67.72	382.55 / 382.55	109	PP-OCRv4 server-side seal text detection model, with higher accuracy, suitable for deployment on better servers
PP-OCRv4_mobile_seal_det	Inference Model/Training Model	96.47	7.82 / 3.09	48.28 / 23.97	4.6	PP-OCRv4 mobile-side seal text detection model, with higher efficiency, suitable for deployment on the edge

Text Recognition Module:

Model	Model Download Links	Recognition Avg Accuracy(%)	GPU Inference Time (ms) [Normal Mode / High-Performance Mode]	CPU Inference Time (ms) [Normal Mode / High-Performance Mode]	Model Storage Size (M)	Introduction
PP-OCRv5_server_rec	Inference Model/Pretrained Model	86.38	8.45/2.36	122.69/122.69	81 M	PP-OCRv5_rec is a next-generation text recognition model. It aims to efficiently and accurately support the recognition of four major languages—Simplified Chinese, Traditional Chinese, English, and Japanese—as well as complex text scenarios such as handwriting, vertical text, pinyin, and rare characters using a single model. While maintaining recognition performance, it balances inference speed and model robustness, providing efficient and accurate technical support for document understanding in various scenarios.
PP-OCRv5_mobile_rec	Inference Model/Pretrained Model	81.29	1.46/5.43	5.32/91.79	16 M
PP-OCRv4_server_rec_doc	Inference Model/Pretrained Model	86.58	6.65 / 2.38	32.92 / 32.92	91 M	PP-OCRv4_server_rec_doc is trained on a mixed dataset of more Chinese document data and PP-OCR training data, building upon PP-OCRv4_server_rec. It enhances the recognition capabilities for some Traditional Chinese characters, Japanese characters, and special symbols, supporting over 15,000 characters. In addition to improving document-related text recognition, it also enhances general text recognition capabilities.
PP-OCRv4_mobile_rec	Inference Model/Pretrained Model	83.28	4.82 / 1.20	16.74 / 4.64	11 M	A lightweight recognition model of PP-OCRv4 with high inference efficiency, suitable for deployment on various hardware devices, including edge devices.
PP-OCRv4_server_rec	Inference Model/Pretrained Model	85.19	6.58 / 2.43	33.17 / 33.17	87 M	The server-side model of PP-OCRv4, offering high inference accuracy and deployable on various servers.
en_PP-OCRv4_mobile_rec	Inference Model/Pretrained Model	70.39	4.81 / 0.75	16.10 / 5.31	7.3 M	An ultra-lightweight English recognition model trained based on the PP-OCRv4 recognition model, supporting English and numeric character recognition.

❗ The above section lists the 6 core models that are primarily supported by the text recognition module. In total, the module supports 20 comprehensive models, including multiple multilingual text recognition models. Below is the complete list of models:

👉Details of the Model List

* PP-OCRv5 Multi-Scenario Models

Model	Model Download Links	Avg Accuracy for Chinese Recognition (%)	Avg Accuracy for English Recognition (%)	Avg Accuracy for Traditional Chinese Recognition (%)	Avg Accuracy for Japanese Recognition (%)	GPU Inference Time (ms) [Normal Mode / High-Performance Mode]	CPU Inference Time (ms) [Normal Mode / High-Performance Mode]	Model Storage Size (M)	Introduction
PP-OCRv5_server_rec	Inference Model/Pretrained Model	86.38	64.70	93.29	60.35	8.45/2.36	122.69/122.69	81 M	PP-OCRv5_rec is a next-generation text recognition model. It aims to efficiently and accurately support the recognition of four major languages—Simplified Chinese, Traditional Chinese, English, and Japanese—as well as complex text scenarios such as handwriting, vertical text, pinyin, and rare characters using a single model. While maintaining recognition performance, it balances inference speed and model robustness, providing efficient and accurate technical support for document understanding in various scenarios.
PP-OCRv5_mobile_rec	Inference Model/Pretrained Model	81.29	66.00	83.55	54.65	1.46/5.43	5.32/91.79	16 M

* Chinese Recognition Model

Model	Model Download Link	Recognition Avg Accuracy(%)	CPU Inference Time (ms) [Normal Mode / High-Performance Mode]	CPU Inference Time (ms) [Normal Mode / High-Performance Mode]	Model Storage Size (M)	Introduction
PP-OCRv4_server_rec_doc	Inference Model/Training Model	81.53	6.65 / 2.38	32.92 / 32.92	74.7 M	PP-OCRv4_server_rec_doc is trained on a mixed dataset of more Chinese document data and PP-OCR training data based on PP-OCRv4_server_rec. It has added the recognition capabilities for some traditional Chinese characters, Japanese, and special characters. The number of recognizable characters is over 15,000. In addition to the improvement in document-related text recognition, it also enhances the general text recognition capability.
PP-OCRv4_mobile_rec	Inference Model/Training Model	78.74	4.82 / 1.20	16.74 / 4.64	10.6 M	The lightweight recognition model of PP-OCRv4 has high inference efficiency and can be deployed on various hardware devices, including edge devices.
PP-OCRv4_server_rec	Inference Model/Training Model	80.61	6.58 / 2.43	33.17 / 33.17	71.2 M	The server-side model of PP-OCRv4 offers high inference accuracy and can be deployed on various types of servers.
PP-OCRv3_mobile_rec	Inference Model/Training Model	72.96	5.87 / 1.19	9.07 / 4.28	9.2 M	PP-OCRv3’s lightweight recognition model is designed for high inference efficiency and can be deployed on a variety of hardware devices, including edge devices.

Model	Model Download Link	Recognition Avg Accuracy(%)	GPU Inference Time (ms) [Normal Mode / High-Performance Mode]	CPU Inference Time (ms) [Normal Mode / High-Performance Mode]	Model Storage Size (M)	Introduction
ch_SVTRv2_rec	Inference Model/Training Model	68.81	8.08 / 2.74	50.17 / 42.50	73.9 M	SVTRv2 is a server text recognition model developed by the OpenOCR team of Fudan University's Visual and Learning Laboratory (FVL). It won the first prize in the PaddleOCR Algorithm Model Challenge - Task One: OCR End-to-End Recognition Task. The end-to-end recognition accuracy on the A list is 6% higher than that of PP-OCRv4.

Model	Model Download Link	Recognition Avg Accuracy(%)	GPU Inference Time (ms) [Normal Mode / High-Performance Mode]	CPU Inference Time (ms) [Normal Mode / High-Performance Mode]	Model Storage Size (M)	Introduction
ch_RepSVTR_rec	Inference Model/Training Model	65.07	5.93 / 1.62	20.73 / 7.32	22.1 M	The RepSVTR text recognition model is a mobile text recognition model based on SVTRv2. It won the first prize in the PaddleOCR Algorithm Model Challenge - Task One: OCR End-to-End Recognition Task. The end-to-end recognition accuracy on the B list is 2.5% higher than that of PP-OCRv4, with the same inference speed.

* English Recognition Model

Model	Model Download Link	Recognition Avg Accuracy(%)	GPU Inference Time (ms) [Normal Mode / High-Performance Mode]	CPU Inference Time (ms) [Normal Mode / High-Performance Mode]	Model Storage Size (M)	Introduction
en_PP-OCRv4_mobile_rec	Inference Model/Training Model	70.39	4.81 / 0.75	16.10 / 5.31	6.8 M	The ultra-lightweight English recognition model trained based on the PP-OCRv4 recognition model supports the recognition of English and numbers.
en_PP-OCRv3_mobile_rec	Inference Model/Training Model	70.69	5.44 / 0.75	8.65 / 5.57	7.8 M	The ultra-lightweight English recognition model trained based on the PP-OCRv3 recognition model supports the recognition of English and numbers.

* Multilingual Recognition Model

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

Test Environment Description:

• Performance Test Environment
  • Test Dataset：
    • Document Image Orientation Classification Module: A self-built dataset using PaddleOCR, covering multiple scenarios such as ID cards and documents, containing 1000 images.
    • Text Image Rectification Model: DocUNet
    • Layout Detection Model: A self-built layout detection dataset using PaddleOCR, including 500 images of common document types such as Chinese and English papers, magazines, contracts, books, exam papers, and research reports.
    • 3-Category Layout Detection Model: A self-built layout detection dataset using PaddleOCR, containing 1154 images of common document types such as Chinese and English papers, magazines, and research reports.
    • 17-Category Region Detection Model: A self-built layout detection dataset using PaddleOCR, including 892 images of common document types such as Chinese and English papers, magazines, and research reports.
    • 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 PaddleOCR, covering various scenarios such as ID cards and documents, containing 1000 images.
    • Seal Text Detection Model: A self-built dataset using PaddleOCR, 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

Before using the seal text recognition production line locally, please ensure that you have completed the installation of the wheel package according to the installation tutorial. Once the installation is complete, you can experience it locally via the command line or integrate it with Python.

2.1 Command Line Experience

You can quickly experience the seal_recognition production line effect with a single command:

paddleocr seal_recognition -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/seal_text_det.png \
    --use_doc_orientation_classify False \
    --use_doc_unwarping False

# Use --device to specify the use of GPU for model inference.
paddleocr seal_recognition -i ./seal_text_det.png --device gpu

The command line supports more parameter settings. Click to expand for detailed explanations of command line parameters.

Parameter	Description	Parameter Type	Default Value
input	Data to be predicted, supporting multiple input types, required. Python Var: Image data represented by numpy.ndarray str: Local path of image or PDF file, e.g., /root/data/img.jpg; URL link, e.g., network URL of image or PDF file: Example; Local directory, the directory should contain images to be predicted, e.g., local path: /root/data/ (currently does not support prediction of PDF files in directories; PDF files must be specified with a specific file path) List: Elements of the list must be 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	Specify the path to save the inference results file. If set to None, the inference results will not be saved locally.	str	None
doc_orientation_classify_model_name	The name of the document orientation classification model. If set to None, the default model in pipeline will be used.	str	None
doc_orientation_classify_model_dir	The directory path of the document orientation classification model. If set to None, the official model will be downloaded.	str	None
doc_unwarping_model_name	The name of the text image unwarping model. If set to None, the default model in pipeline will be used.	str	None
doc_unwarping_model_dir	The directory path of the text image unwarping model. If set to None, the official model will be downloaded.	str	None
layout_detection_model_name	The name of the layout detection model. If set to None, the default model in pipeline will be used.	str	None
layout_detection_model_dir	The directory path of the layout detection model. If set to None, the official model will be downloaded.	str	None
seal_text_detection_model_name	The name of the seal text detection model. If set to None, the production line's default model will be used.	str	None
seal_text_detection_model_dir	The directory path of the seal text detection model. If set to None, the official model will be downloaded.	str	None
text_recognition_model_name	Name of the text recognition model. If None, the default pipeline model is used.	str	None
text_recognition_model_dir	Directory path of the text recognition model. If None, the official model is downloaded.	str	None
text_recognition_batch_size	Batch size for the text recognition model. If None, defaults to 1.	int	None
use_doc_orientation_classify	Whether to enable document orientation classification. If None, defaults to pipeline initialization value (True).	bool	None
use_doc_unwarping	Whether to enable text image correction. If None, defaults to pipeline initialization value (True).	bool	None
use_layout_detection	Whether to load the layout detection module. If set to None, the parameter will default to the value initialized in the pipeline, which is True.	bool	None
layout_threshold	Threshold for layout detection, used to filter out predictions with low confidence. float， such as 0.2, indicates filtering out all bounding boxes with a confidence score less than 0.2. Dictionary, with int keys representing cls_id and float values as thresholds. For example, {0: 0.45, 2: 0.48, 7: 0.4} indicates applying a threshold of 0.45 for class ID 0, 0.48 for class ID 2, and 0.4 for class ID 7 None, If not specified, the default PaddleX official model configuration will be used	float|dict	None
layout_nms	Whether to use NMS (Non-Maximum Suppression) post-processing for layout region detection to filter out overlapping boxes. If set to None, the default configuration of the official model will be used.	bool	None
layout_unclip_ratio	The scaling factor for the side length of the detection boxes in layout region detection. float: A positive float number, e.g., 1.1, indicating that the center of the bounding box remains unchanged while the width and height are both scaled up by a factor of 1.1 List: e.g., [1.2, 1.5], indicating that the center of the bounding box remains unchanged while the width is scaled up by a factor of 1.2 and the height by a factor of 1.5 None: If not specified, the default PaddleX official model configuration will be used	float|list	None
layout_merge_bboxes_mode	The merging mode for the detection boxes output by the model in layout region detection. large: When set to "large", only the largest outer bounding box will be retained for overlapping bounding boxes, and the inner overlapping boxes will be removed. small: When set to "small", only the smallest inner bounding boxes will be retained for overlapping bounding boxes, and the outer overlapping boxes will be removed. union: No filtering of bounding boxes will be performed, and both inner and outer boxes will be retained. None: If not specified, the default PaddleX official model configuration will be used	str	None
seal_det_limit_side_len	The side length limit for seal detection images.	int|None	int: Any integer greater than 0; None: If set to None, it will default to the value initialized by the pipeline, initialized to 960;	None
seal_det_limit_type	The type of side length limit for seal detection images.	str|None	str: Supports min and max, where min ensures that the shortest side of the image is not less than det_limit_side_len, and max ensures that the longest side of the image is not greater than limit_side_len. None: If set to None, it will default to the value initialized by the pipeline, initialized to max;	None
seal_det_thresh	The pixel threshold for detection. In the output probability map, pixel points with scores greater than this threshold will be considered as seal pixels.	float|None	float: Any floating-point number greater than 0. None: If set to None, it will default to the value initialized by the pipeline, initialized to 0.3.	None
seal_det_box_thresh	The bounding box threshold for detection. When the average score of all pixel points within the detection result bounding box is greater than this threshold, the result will be considered as a seal region.	float|None	float: Any floating-point number greater than 0. None: If set to None, it will default to the value initialized by the pipeline, initialized to 0.6.	None
seal_det_unclip_ratio	The expansion coefficient for seal detection. This method is used to expand the seal region, and the larger the value, the larger the expansion area.	float|None	float: Any floating-point number greater than 0. None: If set to None, it will default to the value initialized by the pipeline, initialized to 2.0.	None
seal_rec_score_thresh	The seal recognition threshold. Text results with scores greater than this threshold will be retained.	float|None	float: Any floating-point number greater than 0. None: If set to None, it will default to the value initialized by the pipeline, initialized to 0.0. I.e., no threshold is set.	None
device	The device used for inference. Support for specifying specific card numbers. CPU: For example, cpu indicates using the CPU for inference. GPU: For example, gpu:0 indicates using the first GPU for inference. NPU: For example, npu:0 indicates using the first NPU for inference. XPU: For example, xpu:0 indicates using the first XPU for inference. MLU: For example, mlu:0 indicates using the first MLU for inference. DCU: For example, dcu:0 indicates using the first DCU for inference. None: If set to None, the parameter value initialized by the pipeline will be used by default. During initialization, the local GPU 0 device will be prioritized; if not available, the CPU device 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	The minimum subgraph size, used to optimize the computation of model subgraphs.	int	3
precision	The computational precision, such as fp32, fp16.	str	fp32
enable_mkldnn	Whether to enable the MKL-DNN acceleration library. If set to None, it will be enabled by default.	bool	None
cpu_threads	The number of threads used for inference on the CPU.	int	8
paddlex_config	Path to PaddleX pipeline configuration file.	str	None

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

👉Click to Expand

{'res': {'input_path': 'seal_text_det.png', 'model_settings': {'use_doc_preprocessor': False, 'use_layout_detection': True}, 'layout_det_res': {'input_path': None, 'page_index': None, 'boxes': [{'cls_id': 16, 'label': 'seal', 'score': 0.975531280040741, 'coordinate': [6.195526, 0.1579895, 634.3982, 628.84595]}]}, 'seal_res_list': [{'input_path': None, 'page_index': None, 'model_settings': {'use_doc_preprocessor': False, 'use_textline_orientation': False}, 'dt_polys': [array([[320,  38],
       ...,
       [315,  38]]), array([[461, 347],
       ...,
       [456, 346]]), array([[439, 445],
       ...,
       [434, 444]]), array([[158, 468],
       ...,
       [154, 466]])], 'text_det_params': {'limit_side_len': 736, 'limit_type': 'min', 'thresh': 0.2, 'box_thresh': 0.6, 'unclip_ratio': 0.5}, 'text_type': 'seal', 'textline_orientation_angles': array([-1, ..., -1]), 'text_rec_score_thresh': 0, 'rec_texts': ['天津君和缘商贸有限公司', '发票专用章', '吗繁物', '5263647368706'], 'rec_scores': array([0.9934051 , ..., 0.99139398]), 'rec_polys': [array([[320,  38],
       ...,
       [315,  38]]), array([[461, 347],
       ...,
       [456, 346]]), array([[439, 445],
       ...,
       [434, 444]]), array([[158, 468],
       ...,
       [154, 466]])], 'rec_boxes': array([], dtype=float64)}]}}

The explanation of the result parameters can be found in 2.1.2 Python Script Integration.

The visualized results are saved under save_path, and the visualized result of seal OCR is as follows:

2.2.2 Python Script Integration

• The above command line is for quickly experiencing and viewing the effect. Generally, in a project, you often need to integrate through code. You can complete the quick inference of the pipeline with just a few lines of code. The inference code is as follows:

from paddleocr import SealRecognition

pipeline = SealRecognition(
    use_doc_orientation_classify=False, # Set whether to use document orientation classification model
    use_doc_unwarping=False, # Set whether to use document image unwarping module
)
# ocr = SealRecognition(device="gpu") # Specify GPU for model inference
output = pipeline.predict("./seal_text_det.png")
for res in output:
    res.print() ## Print structured prediction results
    res.save_to_img("./output/")
    res.save_to_json("./output/")

In the above Python script, the following steps were executed:

(1) The seal recognition pipeline object was instantiated via create_pipeline(), with the specific parameters described as follows:

Parameter	Description	Type	Default Value
pipeline	The name of the pipeline or the path to the pipeline configuration file. If it is a pipeline name, it must be supported by PaddleX.	str	None
config	Specific configuration information for the pipeline (if set simultaneously with pipeline, it has higher priority than pipeline, and the pipeline name must be consistent with pipeline).	dict[str, Any]	None
device	The device used for pipeline inference. It supports specifying the specific card number of the GPU, such as "gpu:0", other hardware card numbers, such as "npu:0", or CPU, such as "cpu". Supports specifying multiple devices simultaneously for parallel inference. For details, please refer to Pipeline Parallel Inference.	str	gpu:0
use_hpip	Whether to enable the high-performance inference plugin. If set to None, the setting from the configuration file or config will be used.	bool	None	None
hpi_config	High-performance inference configuration	dict | None	None	None

(2) Call the predict() method of the Seal Text Recognition pipeline object for inference prediction. This method will return a generator. Below are the parameters and their descriptions for the predict() method:

Parameter	Description	Type	Options	Default Value
input	Data to be predicted, supports multiple input types (required)	Python Var|str|list	Python Var: Image data represented by numpy.ndarray str: Local path of an image or PDF file, e.g., /root/data/img.jpg; URL link, e.g., the network URL of an image or PDF file: Example; Local directory, containing images to be predicted, e.g., /root/data/ (currently does not support prediction of PDF files in directories; PDF files must be specified with an exact file path) List: Elements of the list must be of the above types, e.g., [numpy.ndarray, numpy.ndarray], [\"/root/data/img1.jpg\", \"/root/data/img2.jpg\"], [\"/root/data1\", \"/root/data2\"]
device	Inference device for the pipeline	str|None	CPU: e.g., cpu for CPU inference; GPU: e.g., gpu:0 for inference using the first GPU; NPU: e.g., npu:0 for inference using the first NPU; XPU: e.g., xpu:0 for inference using the first XPU; MLU: e.g., mlu:0 for inference using the first MLU; DCU: e.g., dcu:0 for inference using the first DCU; None: If set to None, the default value from the pipeline initialization will be used. During initialization, the local GPU device 0 will be prioritized; if unavailable, the CPU device will be used.	None
use_doc_orientation_classify	Whether to use the document orientation classification module	bool|None	bool: True or False; None: If set to None, the default value from the pipeline initialization will be used, initialized as True.	None
use_doc_unwarping	Whether to use the document unwarping module	bool|None	bool: True or False; None: If set to None, the default value from the pipeline initialization will be used, initialized as True.	None
use_layout_detection	Whether to use the layout detection module	bool|None	bool: True or False; None: If set to None, the default value from the pipeline initialization will be used, initialized as True.	None
layout_threshold	Confidence threshold for layout detection; only scores above this threshold will be output	float|dict|None	float: Any float greater than 0 dict: Key is the int category ID, value is any float greater than 0 None: If set to None, the default value from the pipeline initialization will be used, initialized as 0.5	None
layout_nms	Whether to use Non-Maximum Suppression (NMS) for layout detection post-processing	bool|None	bool: True or False; None: If set to None, the default value from the pipeline initialization will be used, initialized as True.	None
layout_unclip_ratio	Expansion ratio of detection box edges; if not specified, the default value from the PaddleX official model configuration will be used	float|list|None	float: Any float greater than 0, e.g., 1.1, which means expanding the width and height of the detection box by 1.1 times while keeping the center unchanged list: e.g., [1.2, 1.5], which means expanding the width of the detection box by 1.2 times and the height by 1.5 times while keeping the center unchanged None: If set to None, the default value from the pipeline initialization will be used, initialized as 1.0
layout_merge_bboxes_mode	Merging mode for detection boxes in layout detection output; if not specified, the default value from the PaddleX official model configuration will be used	string|None	large: When set to large, only the largest external box will be retained for overlapping detection boxes, and the internal overlapping boxes will be removed. small: When set to small, only the smallest internal box will be retained for overlapping detection boxes, and the external overlapping boxes will be removed. union: No filtering of boxes will be performed; both internal and external boxes will be retained. None: If set to None, the default value from the pipeline initialization will be used, initialized as large.	None
seal_det_limit_side_len	Side length limit for seal text detection	int|None	int: Any integer greater than 0 None: If set to None, the default value from the pipeline initialization will be used, initialized as 736	None
seal_rec_score_thresh	Text recognition threshold; text results with scores above this threshold will be retained	float|None	float: Any float greater than 0 None: If set to None, the default value from the pipeline initialization will be used, initialized as 0.0. This means no threshold is applied.	None

(3) Process the prediction results. The prediction result for each sample is of dict type and supports operations such as printing, saving as an image, and saving as a json file:

Method	Description	Parameter	Parameter Type	Parameter Description	Default Value
print()	Print results to the terminal	format_json	bool	Whether to format the output content using JSON indentation	True
		indent	int	Specify the indentation level to beautify the output JSON data for better readability, effective only when format_json is True	4
		ensure_ascii	bool	Control whether to escape non-ASCII characters to Unicode. When set to True, all non-ASCII characters will be escaped; False will retain the original characters, effective only when format_json is True	False
save_to_json()	Save results as a json file	save_path	str	The file path to save the results. When it is a directory, the saved file name will be consistent with the input file type	None
		indent	int	Specify the indentation level to beautify the output JSON data for better readability, effective only when format_json is True	4
		ensure_ascii	bool	Control whether to escape non-ASCII characters to Unicode. When set to True, all non-ASCII characters will be escaped; False will retain the original characters, effective only when format_json is True	False
save_to_img()	Save results as an image file	save_path	str	The file path to save the results, supports directory or file path	None

• Calling the print() method will print the results to the terminal, and the explanations of the printed content are as follows:

  • input_path: (str) The input path of the image to be predicted.

  • model_settings: (Dict[str, bool]) The model parameters required for pipeline configuration.

    • use_doc_preprocessor: (bool) Controls whether to enable the document preprocessing sub-pipeline.
    • use_layout_detection: (bool) Controls whether to enable the layout detection sub-module.

  • layout_det_res: (Dict[str, Union[List[numpy.ndarray], List[float]]]) The output result of the layout detection sub-module. Only exists when use_layout_detection=True.

    • input_path: (Union[str, None]) The image path accepted by the layout detection module. Saved as None when the input is a numpy.ndarray.
    • page_index: (Union[int, None]) Indicates the current page number of the PDF if the input is a PDF file; otherwise, it is None.
    • boxes: (List[Dict]) A list of detected layout seal regions, with each element containing the following fields:
      • cls_id: (int) The class ID of the detected seal region.
      • score: (float) The confidence score of the detected region.
      • coordinate: (List[float]) The coordinates of the four corners of the detection box, in the order of x1, y1, x2, y2, representing the x-coordinate of the top-left corner, the y-coordinate of the top-left corner, the x-coordinate of the bottom-right corner, and the y-coordinate of the bottom-right corner.

  • seal_res_list: List[Dict] A list of seal text recognition results, with each element containing the following fields:

    • input_path: (Union[str, None]) The image path accepted by the seal text recognition pipeline. Saved as None when the input is a numpy.ndarray.
    • page_index: (Union[int, None]) Indicates the current page number of the PDF if the input is a PDF file; otherwise, it is None.
    • model_settings: (Dict[str, bool]) The model configuration parameters for the seal text recognition pipeline.
    • use_doc_preprocessor: (bool) Controls whether to enable the document preprocessing sub-pipeline.
    • use_textline_orientation: (bool) Controls whether to enable the text line orientation classification sub-module.

  • doc_preprocessor_res: (Dict[str, Union[str, Dict[str, bool], int]]) The output result of the document preprocessing sub-pipeline. Only exists when use_doc_preprocessor=True.

    • input_path: (Union[str, None]) The image path accepted by the document preprocessing sub-pipeline. Saved as None when the input is a numpy.ndarray.
    • model_settings: (Dict) The model configuration parameters for the preprocessing sub-pipeline.
      • use_doc_orientation_classify: (bool) Controls whether to enable document orientation classification.
      • use_doc_unwarping: (bool) Controls whether to enable document unwarping.
    • angle: (int) The predicted result of document orientation classification. When enabled, it takes values [0, 1, 2, 3], corresponding to [0°, 90°, 180°, 270°]; when disabled, it is -1.

  • dt_polys: (List[numpy.ndarray]) A list of polygon boxes for seal text detection. Each detection box is represented by a numpy array of multiple vertex coordinates, with the array shape being (n, 2).

  • dt_scores: (List[float]) A list of confidence scores for text detection boxes.

  • text_det_params: (Dict[str, Dict[str, int, float]]) Configuration parameters for the text detection module.

    • limit_side_len: (int) The side length limit value during image preprocessing.
    • limit_type: (str) The handling method for side length limits.
    • thresh: (float) The confidence threshold for text pixel classification.
    • box_thresh: (float) The confidence threshold for text detection boxes.
    • unclip_ratio: (float) The expansion ratio for text detection boxes.
    • text_type: (str) The type of seal text detection, currently fixed as "seal".

  • text_rec_score_thresh: (float) The filtering threshold for text recognition results.

  • rec_texts: (List[str]) A list of text recognition results, containing only texts with confidence scores above text_rec_score_thresh.

  • rec_scores: (List[float]) A list of confidence scores for text recognition, filtered by text_rec_score_thresh.

  • rec_polys: (List[numpy.ndarray]) A list of text detection boxes filtered by confidence score, in the same format as dt_polys.

  • rec_boxes: (numpy.ndarray) An array of rectangular bounding boxes for detection boxes; the seal recognition pipeline returns an empty array.

• Calling the save_to_json() method will save the above content to the specified save_path. If a directory is specified, the saved path will be save_path/{your_img_basename}_res.json. If a file is specified, it will be saved directly to that file. Since JSON files do not support saving numpy arrays, numpy.array types will be converted to list format.

• Calling the save_to_img() method will save the visualization results to the specified save_path. If a directory is specified, the saved path will be save_path/{your_img_basename}_seal_res_region1.{your_img_extension}. If a file is specified, it will be saved directly to that file. (The pipeline usually contains multiple result images, so it is not recommended to specify a specific file path directly, as multiple images will be overwritten, and only the last image will be retained.)

• Additionally, you can obtain visualized images with results and prediction results through attributes, as follows:

Attribute	Description
json	Get the prediction results in json format.
img	Get the visualization results in dict format.

• The prediction results obtained through the json attribute are of dict type, with content consistent with what is saved by calling the save_to_json() method.
• The prediction results returned by the img attribute are of dict type. The keys are layout_det_res, seal_res_region1, and preprocessed_img, corresponding to three Image.Image objects: one for visualizing layout detection, one for visualizing seal text recognition results, and one for visualizing image preprocessing. If the image preprocessing sub-module is not used, preprocessed_img will not be included in the dictionary. If the layout region detection module is not used, layout_det_res will not be included.

3. Development Integration/Deployment

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

If you need to integrate the pipeline into your Python project, you can refer to the example code in 2.2 Python Script Method.

In addition, PaddleX also provides three other deployment methods, which are detailed as follows:

🚀 High-Performance Inference: In real-world production environments, many applications have stringent performance requirements for deployment strategies, especially in terms of response speed, to ensure efficient system operation and a smooth user experience. To address this, PaddleOCR offers high-performance inference capabilities aimed at deeply optimizing the performance of model inference and pre/post-processing, thereby significantly accelerating the end-to-end process. For detailed high-performance inference procedures, please refer to High-Performance Inference.

☁️ Service Deployment: Service deployment is a common form of deployment in real-world production environments. By encapsulating inference functionality into a service, clients can access these services via network requests to obtain inference results. For detailed production service deployment procedures, please refer to Serving.

Below are the API references for basic serving deployment and multi-language service invocation examples:

API Reference

For the main operations provided by the service:

• The HTTP request method is POST.
• The request body and response body are both JSON data (JSON objects).
• When the request is processed successfully, the response status code is 200, and the attributes of the response body are as follows:

Name	Type	Description
logId	string	The UUID of the request.
errorCode	integer	Error code. Fixed as 0.
errorMsg	string	Error message. Fixed as "Success".
result	object	The result of the operation.

• When the request is not processed successfully, the attributes of the response body are as follows:

Name	Type	Description
logId	string	The UUID of the request.
errorCode	integer	Error code. Same as the response status code.
errorMsg	string	Error message.

The main operations provided by the service are as follows:

• infer

Obtain the seal text recognition result.

POST /seal-recognition

• The attributes of the request body are as follows:

Name	Type	Description	Required
file	string	The URL of an image or PDF file accessible by the server, or the Base64-encoded content of the file. By default, for PDF files exceeding 10 pages, only the content of the first 10 pages will be processed. To remove the page limit, please add the following configuration to the pipeline configuration file: Serving: extra: max_num_input_imgs: null	Yes
fileType	integer | null	The type of file. 0 indicates a PDF file, 1 indicates an image file. If this attribute is not present in the request body, the file type will be inferred from the URL.	No
useDocOrientationClassify	boolean | null	Please refer to the description of the use_doc_orientation_classify parameter of the pipeline object's predict method.	No
useDocUnwarping	boolean | null	Please refer to the description of the use_doc_unwarping parameter of the pipeline object's predict method.	No
useLayoutDetection	boolean | null	Please refer to the description of the use_layout_detection parameter of the pipeline object's predict method.	No
layoutThreshold	number | null	Please refer to the description of the layout_threshold parameter of the pipeline object's predict method.	No
layoutNms	boolean | null	Please refer to the description of the layout_nms parameter of the pipeline object's predict method.	No
layoutUnclipRatio	number | array | null	Please refer to the description of the layout_unclip_ratio parameter of the pipeline object's predict method.	No
layoutMergeBboxesMode	string | null	Please refer to the description of the layout_merge_bboxes_mode parameter of the pipeline object's predict method.	No
sealDetLimitSideLen	integer | null	Please refer to the description of the seal_det_limit_side_len parameter of the pipeline object's predict method.	No
sealDetLimitType	string | null	Please refer to the description of the seal_det_limit_type parameter of the pipeline object's predict method.	No
sealDetThresh	number | null	Please refer to the description of the seal_det_thresh parameter of the pipeline object's predict method.	No
sealDetBoxThresh	number | null	Please refer to the description of the seal_det_box_thresh parameter of the pipeline object's predict method.	No
sealDetUnclipRatio	number | null	Please refer to the description of the seal_det_unclip_ratio parameter of the pipeline object's predict method.	No
sealRecScoreThresh	number | null	Please refer to the description of the seal_rec_score_thresh parameter of the pipeline object's predict method.	No

• When the request is processed successfully, the result in the response body has the following properties:

Name	Type	Meaning
sealRecResults	object	The seal text recognition result. The array length is 1 (for image input) or the actual number of document pages processed (for PDF input). For PDF input, each element in the array represents the result of each page actually processed in the PDF file.
dataInfo	object	Information about the input data.

Each element in sealRecResults is an object with the following properties:

Name	Type	Meaning
prunedResult	object	A simplified version of the res field in the JSON representation generated by the predict method of the production object, where the input_path and the page_index fields are removed.
outputImages	object | null	See the description of the img attribute of the result of the pipeline prediction. The images are in JPEG format and encoded in Base64.
inputImage	string | null	The input image. The image is in JPEG format and encoded in Base64.

Multi-language Service Invocation Example

Python

import base64
import requests

API_URL = "http://localhost:8080/seal-recognition"
file_path = "./demo.jpg"

with open(file_path, "rb") as file:
    file_bytes = file.read()
    file_data = base64.b64encode(file_bytes).decode("ascii")

payload = {"file": file_data, "fileType": 1}

response = requests.post(API_URL, json=payload)

assert response.status_code == 200
result = response.json()["result"]
for i, res in enumerate(result["sealRecResults"]):
    print(res["prunedResult"])
    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. Custom Development

If the default model weights provided by the seal text recognition pipeline do not meet your requirements in terms of accuracy or speed, you can try to fine-tune the existing models using your own domain-specific or application data to improve the recognition performance of the seal text recognition pipeline in your scenario.

Since the seal text recognition pipeline consists of several modules, if the pipeline's performance does not meet expectations, the issue may arise from any one of these modules. You can analyze images with poor recognition results to identify which module is problematic and refer to the corresponding fine-tuning tutorial links in the table below for model fine-tuning.

Scenario	Fine-Tuning Module	Fine-Tuning Reference Link
Inaccurate or missing seal position detection	Layout Detection Module	Link
Missing text detection	Text Detection Module	Link
Inaccurate text content	Text Recognition Module	Link
Inaccurate full-image rotation correction	Document Image Orientation Classification Module	Link
Inaccurate image distortion correction	Text Image Correction Module	Not supported for fine-tuning

Comments