Formula Recognition Pipeline User Guide¶

1. Introduction to Formula Recognition Pipeline¶

Formula recognition is a technology that automatically identifies and extracts LaTeX formula content and structure from documents or images. It is widely used in fields such as mathematics, physics, and computer science for document editing and data analysis. By using computer vision and machine learning algorithms, formula recognition can convert complex mathematical formula information into editable LaTeX format, facilitating further processing and analysis of data.

The formula recognition pipeline is designed to solve formula recognition tasks by extracting formula information from images and outputting it in LaTeX source code format. This pipeline integrates the advanced formula recognition model PP-FormulaNet developed by the PaddlePaddle Vision Team and the well-known formula recognition model UniMERNet. It is an end-to-end formula recognition system that supports the recognition of simple printed formulas, complex printed formulas, and handwritten formulas. Additionally, it includes functions for image orientation correction and distortion correction. Based on this pipeline, precise formula content prediction can be achieved, covering various application scenarios in education, research, finance, manufacturing, and other fields. The pipeline also provides flexible deployment options, supporting multiple hardware devices and programming languages. Moreover, it offers the capability for secondary development. You can train and optimize the pipeline on your own dataset, and the trained model can be seamlessly integrated.

The formula recognition pipeline includes a mandatory formula recognition module, as well as optional layout detection, document image orientation classification, and text image unwarping modules. The document image orientation classification module and the text image unwarping module are integrated into the formula recognition pipeline as a document preprocessing sub-pipeline. Each module contains multiple models, and you can choose the model based on the benchmark test data below.

If you prioritize model accuracy, choose a model with higher precision; if you care more about inference speed, choose a faster model; if you are concerned about model storage size, choose a smaller model.

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)	Introduction
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, with 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)	Introduction
UVDoc	Inference Model/Training Model	0.179	30.3 M	High-precision text image correction model

Layout Detection Module (Optional):

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 3 core models that are key supported by the text recognition module. The module actually supports a total of 6 full models, including several predefined models with different categories. The complete model list is as follows:

👉 Details of Model List

* 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.

* 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.

Formula Recognition Module

Model	Model Download Link	Avg-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	86.13	2266.96/-	-/-	1.4 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	87.12	202.25/-	-/-	167.9 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	92.13	1976.52/-	-/-	535.2 M
LaTeX_OCR_rec	Inference Model/Training Model	71.63	-/-	-/-	89.7 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.

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 Module: DocUNet.
- Layout Region Detection Module: A self-built layout region 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.
- 17-Class Region Detection Model: A self-built layout region detection dataset using PaddleOCR, including 892 images of common document types such as Chinese and English papers, magazines, and research reports.
- Formula Recognition Module: A self-built formula recognition test set using PaddleX.
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 model pipelines provided by PaddleX can be quickly experienced. You can experience the effect of the formula recognition pipeline on the community platform, or you can use the command line or Python locally to experience the effect of the formula recognition pipeline.

2.1 Online Experience¶

You can experience the formula recognition pipeline online by recognizing the demo images provided by the official platform, for example:

If you are satisfied with the performance of the pipeline, you can directly integrate and deploy it. You can choose to download the deployment package from the cloud, or refer to the methods in Section 2.2 Local Experience for local deployment. If you are not satisfied with the effect, you can fine-tune the models in the pipeline using your private data. If you have local hardware resources for training, you can start training directly on your local machine; if not, the Star River Zero-Code platform provides a one-click training service. You don't need to write any code—just upload your data and start the training task with one click.

2.2 Local Experience¶

❗ Before using the formula recognition pipeline locally, please ensure that you have completed the installation of the PaddleX wheel package according to the PaddleX Installation Guide.

2.2.1 Command Line Experience¶

You can quickly experience the effect of the formula recognition pipeline with one command. Use the test file, and replace --input with the local path for prediction.

paddlex --pipeline formula_recognition \
        --input general_formula_recognition_001.png \
        --use_layout_detection True \
        --use_doc_orientation_classify False \
        --use_doc_unwarping False \
        --layout_threshold 0.5 \
        --layout_nms True \
        --layout_unclip_ratio  1.0 \
        --layout_merge_bboxes_mode large \
        --save_path ./output \
        --device gpu:0

The relevant parameter descriptions can be referenced from 2.2 Integration via Python Script.

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

{'res': {'input_path': 'general_formula_recognition.png', 'page_index': None, 'model_settings': {'use_doc_preprocessor': False,'use_layout_detection': True}, 'layout_det_res': {'input_path': None, 'boxes': [{'cls_id': 2, 'label': 'text', 'score': 0.9778407216072083, 'coordinate': [271.257, 648.50824, 1040.2291, 774.8482]}, ...]}, 'formula_res_list': [{'rec_formula': '\\small\\begin{aligned}{p(\\mathbf{x})=c(\\mathbf{u})\\prod_{i}p(x_{i}).}\\\\ \\end{aligned}', 'formula_region_id': 1, 'dt_polys': ([553.0718, 802.0996, 758.75635, 853.093],)}, ...]}}

The explanation of the running result parameters can refer to the result interpretation in 2.2 Integration via Python Script.

The visualization results are saved under save_path, where the visualization result of formula recognition is as follows:

If you need to visualize the formula recognition pipeline, you need to run the following command to install the LaTeX rendering environment. Currently, visualization of the formula recognition pipeline only supports the Ubuntu environment, and other environments are not supported. For complex formulas, the LaTeX result may contain some advanced representations that may not be successfully displayed in environments such as Markdown:

sudo apt-get update
sudo apt-get install texlive texlive-latex-base texlive-latex-extra -y

Note: Due to the need to render each formula image during the formula recognition visualization process, the process takes a long time. Please be patient.

2.2.2 Python Script Integration¶

A few lines of code can quickly complete the pipeline inference. Taking the formula recognition pipeline as an example:

from paddlex import create_pipeline

pipeline = create_pipeline(pipeline="formula_recognition")

output = pipeline.predict(
    input="./general_formula_recognition_001.png",
    use_layout_detection=True ,
    use_doc_orientation_classify=False,
    use_doc_unwarping=False,
    layout_threshold=0.5,
    layout_nms=True,
    layout_unclip_ratio=1.0,
    layout_merge_bboxes_mode="large"
)
for res in output:
    res.print()
    res.save_to_img(save_path="./output/")
    res.save_to_json(save_path="./output/")

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

(1) Instantiate the formula recognition pipeline object through create_pipeline(), with specific parameters as follows:

Parameter	Description	Type	Default
`pipeline`	Pipeline name or path to pipeline config file, if it's set as a pipeline name, it must be a pipeline supported by PaddleX.	`str`	`None`
`config`	Specific configuration information for the pipeline (if set simultaneously with the `pipeline`, it takes precedence over the `pipeline`, and the pipeline name must match the `pipeline`).	`dict[str, Any]`	`None`
`device`	Pipeline inference device. Supports specifying the specific GPU card number, such as "gpu:0", other hardware specific card numbers, such as "npu:0", CPU such as "cpu".	`str`	`None`
`use_hpip`	Whether to enable high-performance inference, only available when the pipeline supports high-performance inference.	`bool`	`False`

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

Parameter	Description	Type	Options	Default Value
`input`	Data to be predicted, supporting multiple input types, required	`Python Var\|str\|list`	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\"]`	`None`
`device`	pipeline inference device	`str\|None`	CPU: e.g., `cpu` indicates using CPU for inference; GPU: e.g., `gpu:0` indicates using the 1st GPU for inference; NPU: e.g., `npu:0` indicates using the 1st NPU for inference; XPU: e.g., `xpu:0` indicates using the 1st XPU for inference; MLU: e.g., `mlu:0` indicates using the 1st MLU for inference; DCU: e.g., `dcu:0` indicates using the 1st DCU for inference; None: If set to `None`, the default value initialized by the pipeline will be used. During initialization, the local GPU 0 will be prioritized; if unavailable, the CPU will be used.	`None`
`use_layout_detection`	Whether to use the document layout detection module	`bool\|None`	bool: `True` or `False`; None: If set to `None`, the default value initialized by the pipeline will be used, initialized as `True`.	`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 initialized by the pipeline 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 initialized by the pipeline will be used, initialized as `True`.	`None`
`layout_threshold`	Threshold for filtering low-confidence prediction results; if not specified, the default PaddleX official model configuration will be used	`float/dict/None`	float, e.g., 0.2, indicating filtering out all bounding boxes with confidence scores below 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	`None`
`layout_nms`	Whether to use NMS post-processing to filter overlapping bounding boxes; if not specified, the default PaddleX official model configuration will be used	`bool/None`	bool: `True` or `False`, indicating whether to use NMS for post-processing to filter overlapping bounding boxes None: If not specified, the default PaddleX official model configuration will be used	`None`
`layout_unclip_ratio`	Scaling factor for the side length of bounding boxes; if not specified, the default PaddleX official model configuration will be used	`float/list/None`	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	`None`
`layout_merge_bboxes_mode`	Merging mode for the bounding boxes output by the model; if not specified, the default PaddleX official model configuration will be used	`string/None`	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	`None`

(3) Process the prediction results. The prediction result of 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 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, making it more readable. 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` retains the original characters. Effective only when `format_json` is `True`	`False`
`save_to_json()`	Save results as a JSON file	`save_path`	`str`	Path to save the file. If it is a directory, the saved file will be named the same as the input file type	None
		`indent`	`int`	Specify the indentation level to beautify the output `JSON` data, making it more readable. 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` retains the original characters. Effective only when `format_json` is `True`	`False`
`save_to_img()`	Save results as an image file	`save_path`	`str`	Path to save the file, supports directory or file path	None

Calling the print() method will print the results to the terminal. The content printed to the terminal is explained as follows:
- input_path: (str) The input path of the image to be predicted.
- page_index: (Union[int, None]) If the input is a PDF file, this indicates the current page number of the PDF. Otherwise, it is None
- model_settings: (Dict[str, bool]) The model parameters required for the 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 area detection module.
- doc_preprocessor_res: (Dict[str, Union[str, Dict[str, bool], int]]) The output result of the document preprocessing sub-pipeline. It exists only when use_doc_preprocessor=True.
  - input_path: (Union[str, None]) The image path accepted by the image preprocessing sub-pipeline. When the input is a numpy.ndarray, it is saved as None.
  - model_settings: (Dict) The model configuration parameters of 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 distortion correction.
  - angle: (int) The prediction result of document orientation classification. When enabled, it takes values from [0,1,2,3], corresponding to [0°,90°,180°,270°]; when disabled, it is -1.
- layout_det_res: (Dict[str, List[Dict]]) The output result of the layout area detection module. It exists only when use_layout_detection=True.
  - input_path: (Union[str, None]) The image path accepted by the layout area detection module. When the input is a numpy.ndarray, it is saved as None.
  - boxes: (List[Dict[int, str, float, List[float]]]) A list of layout area detection prediction results.
    - cls_id: (int) The class ID predicted by layout area detection.
    - label: (str) The class label predicted by layout area detection.
    - score: (float) The confidence score of the predicted class.
    - coordinate: (List[float]) The bounding box coordinates predicted by layout area detection, in the format [x_min, y_min, x_max, y_max], where (x_min, y_min) is the top-left corner and (x_max, y_max) is the bottom-right corner.
- formula_res_list: (List[Dict[str, int, List[float]]]) A list of formula recognition prediction results.
  - rec_formula: (str) The LaTeX source code predicted by formula recognition.
  - formula_region_id: (int) The ID number predicted by formula recognition.
  - dt_polys: (List[float]) The bounding box coordinates predicted by formula recognition, in the format [x_min, y_min, x_max, y_max], where (x_min, y_min) is the top-left corner and (x_max, y_max) is the bottom-right corner.
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}_formula_res_img.{your_img_extension}. If a file is specified, it will be saved directly to that file. (The pipeline usually contains many result images, so it is not recommended to specify a specific file path directly, otherwise multiple images will be overwritten and only the last one will be retained.)
In addition, you can also obtain the visualization image with results and the prediction results through attributes, as follows:

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

The prediction results obtained through the json attribute are of the dict type, with content consistent with what is saved using the save_to_json() method.
The prediction results returned by the img attribute are of the dictionary type. The keys are preprocessed_img, layout_det_res, and formula_res_img, corresponding to three Image.Image objects: the first one displays the visualization image of image preprocessing, the second one displays the visualization image of layout area detection, and the third one displays the visualization image of formula recognition. If the image preprocessing sub-module is not used, the dictionary does not contain preprocessed_img; if the layout area detection sub-module is not used, the dictionary does not contain layout_det_res.

In addition, you can obtain the configuration file of the formula recognition pipeline and load the configuration file for prediction. You can execute the following command to save the results in my_path:

paddlex --get_pipeline_config formula_recognition --save_path ./my_path

If you have obtained the configuration file, you can customize the settings for the formula recognition pipeline by simply modifying the value of the pipeline parameter in the create_pipeline method to the path of the pipeline configuration file. An example is shown below:

from paddlex import create_pipeline

pipeline = create_pipeline(pipeline="./my_path/formula_recognition.yaml")

output = pipeline.predict(
    input="./general_formula_recognition_001.png",
    use_layout_detection=True ,
    use_doc_orientation_classify=False,
    use_doc_unwarping=False,
    layout_threshold=0.5,
    layout_nms=True,
    layout_unclip_ratio=1.0,
    layout_merge_bboxes_mode="large"
)
for res in output:
    res.print()
    res.save_to_img(save_path="./output/")
    res.save_to_json(save_path="./output/")

Note: The parameters in the configuration file are initialization parameters for the pipeline. If you want to change the initialization parameters for the formula recognition pipeline, you can directly modify the parameters in the configuration file and load the configuration file for prediction. Additionally, CLI prediction also supports passing in a configuration file, simply specify the path of the configuration file with --pipeline.

3. Development Integration/Deployment¶

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

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

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

🚀 High-Performance Inference: In actual production environments, many applications have strict performance requirements for deployment strategies, especially in terms of response speed, to ensure efficient system operation and smooth user experience. To this end, PaddleX provides a high-performance inference plugin, which aims to deeply optimize the performance of model inference and pre/post-processing, significantly speeding up the end-to-end process. For detailed high-performance inference procedures, please refer to the PaddleX High-Performance Inference Guide.

☁️ Service-Based Deployment: Service-based deployment is a common form of deployment in actual production environments. By encapsulating inference capabilities into services, clients can access these services via network requests to obtain inference results. PaddleX supports multiple pipeline service-based deployment solutions. For detailed pipeline service-based deployment procedures, please refer to the PaddleX Service-Based Deployment Guide.

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

API Reference

For the main operations provided by the service:

The HTTP request method is POST.
Both the request body and response body are 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	Meaning
`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	Meaning
`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 formula recognition results from images.

POST /formula-recognition

The attributes of the request body are as follows:

Name	Type	Meaning	Required
`file`	`string`	The URL of an image or PDF file accessible by the server, or the Base64-encoded content of the file. For PDF files exceeding 10 pages, only the first 10 pages will be used.	Yes
`fileType`	`integer` \| `null`	The type of the file. `0` for PDF files, `1` for image files. If this attribute is missing, the file type will be inferred from the URL.	No
`useDocOrientationClassify`	`boolean` \| `null`	Refer to the `use_doc_orientation_classify` parameter description in the pipeline `predict` method.	No
`useDocUnwarping`	`boolean` \| `null`	Refer to the `use_doc_unwarping` parameter description in the pipeline `predict` method.	No
`useLayoutDetection`	`boolean` \| `null`	Refer to the `use_layout_detection` parameter description in the pipeline `predict` method.	No

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

Name	Type	Meaning
`formulaRecResults`	`object`	The formula recognition results. The array length is 1 (for image input) or the smaller of the number of document pages and 10 (for PDF input). For PDF input, each element in the array represents the processing result of each page in the PDF file.
`dataInfo`	`object`	Information about the input data.

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

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

Multi-language Service Invocation Example

Python

import base64
import requests

API_URL = "http://localhost:8080/formula-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["formulaRecResults"]):
    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}")

📱 Edge Deployment: Edge deployment is a method that places computing and data processing capabilities directly on user devices, allowing the device to process data without relying on remote servers. PaddleX supports deploying models on edge devices such as Android. For detailed deployment procedures, please refer to the PaddleX Edge Deployment Guide. You can choose the appropriate deployment method based on your needs to integrate the model pipeline into subsequent AI applications.

4. Secondary Development¶

If the default model weights provided by the formula 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-specific data to improve the recognition performance of the formula recognition pipeline in your scenario.

4.1 Model Fine-Tuning¶

Since the formula recognition pipeline consists of several modules, if the pipeline's performance is not satisfactory, the issue may arise from any one of these modules. You can analyze the poorly recognized images to determine 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	Reference Link
Formulas are missing	Layout Detection Module	Link
Formula content is inaccurate	Formula Recognition Module	Link
Whole-image rotation correction is inaccurate	Document Image Orientation Classification Module	Link
Image distortion correction is inaccurate	Text Image Correction Module	Fine-tuning not supported

4.2 Model Application¶

After fine-tuning with your private dataset, you will obtain the local model weight file.

If you need to use the fine-tuned model weights, simply modify the pipeline configuration file and replace the local path of the fine-tuned model weights into the corresponding position in the pipeline configuration file:

...
SubModules:
  LayoutDetection:
    module_name: layout_detection
    model_name: PP-DocLayout-L
    model_dir: null # Replace with the fine-tuned layout detection model weights path
...

  FormulaRecognition:
    module_name: formula_recognition
    model_name: PP-FormulaNet-L
    model_dir: null # Replace with the fine-tuned formula recognition model weights path
    batch_size: 5

SubPipelines:
  DocPreprocessor:
    pipeline_name: doc_preprocessor
    use_doc_orientation_classify: True
    use_doc_unwarping: True
    SubModules:
      DocOrientationClassify:
        module_name: doc_text_orientation
        model_name: PP-LCNet_x1_0_doc_ori
        model_dir: null # Replace with the fine-tuned document image orientation classification model weights path
        batch_size: 1
...

Then, refer to the command-line or Python script methods in 2. Quick Start to load the modified pipeline configuration file.

5. Multi-Hardware Support¶

PaddleX supports a variety of mainstream hardware devices, including NVIDIA GPU, Kunlunxin XPU, Ascend NPU, and Cambricon MLU. You can seamlessly switch between different hardware devices by simply modifying the --device parameter.

For example, if you use Ascend NPU for formula recognition pipeline inference, the CLI command is:

paddlex --pipeline formula_recognition \
        --input general_formula_recognition_001.png \
        --use_layout_detection True \
        --use_doc_orientation_classify False \
        --use_doc_unwarping False \
        --layout_threshold 0.5 \
        --layout_nms True \
        --layout_unclip_ratio  1.0 \
        --layout_merge_bboxes_mode large \
        --save_path ./output \
        --device npu:0

Of course, you can also specify the hardware device when calling create_pipeline() or predict() in a Python script.

If you want to use the formula recognition pipeline on more types of hardware, please refer to the PaddleX Multi-Hardware Usage Guide.