PGNet
1. Brief Introduction¶
OCR algorithms can be divided into two categories: two-stage algorithm and end-to-end algorithm. The two-stage OCR algorithm is generally divided into two parts, text detection and text recognition algorithm. The text detection algorithm locates the box of the text line from the image, and then the recognition algorithm identifies the content of the text box. The end-to-end OCR algorithm combines text detection and recognition in one algorithm. Its basic idea is to design a model with both detection unit and recognition module, share the CNN features of both and train them together. Because one algorithm can complete character recognition, the end-to-end model is smaller and faster.
Introduction Of PGNet Algorithm¶
During the recent years, the end-to-end OCR algorithm has been well developed, including MaskTextSpotter series, TextSnake, TextDragon, PGNet series and so on. Among these algorithms, PGNet algorithm has some advantages over the other algorithms.
- PGNet loss is designed to guide training, and no character-level annotations is needed.
- NMS and ROI related operations are not needed. It can accelerate the prediction
- The reading order prediction module is proposed
- A graph based modification module (GRM) is proposed to further improve the performance of model recognition
- Higher accuracy and faster prediction speed
For details of PGNet algorithm, please refer to paper. The schematic diagram of the algorithm is as follows:
After feature extraction, the input image is sent to four branches: TBO module for text edge offset prediction, TCL module for text center-line prediction, TDO module for text direction offset prediction, and TCC module for text character classification graph prediction. The output of TBO and TCL can get text detection results after post-processing, and TCL, TDO and TCC are responsible for text recognition.
The results of detection and recognition are as follows:
Performance¶
Test set: Total Text¶
Test environment: NVIDIA Tesla V100-SXM2-16GB¶
| PGNetA | det_precision | det_recall | det_f_score | e2e_precision | e2e_recall | e2e_f_score | FPS | download |
|---|---|---|---|---|---|---|---|---|
| Paper | 85.30 | 86.80 | 86.10 | - | - | 61.70 | 38.20 (size=640) | - |
| Ours | 87.03 | 82.48 | 84.69 | 61.71 | 58.43 | 60.03 | 48.73 (size=768) | download link |
note:PGNet in PaddleOCR optimizes the prediction speed, and can significantly improve the end-to-end prediction speed within the acceptable range of accuracy reduction
2. Environment Configuration¶
Please refer to Operation Environment Preparation to configure PaddleOCR operating environment first, refer to Project Clone to clone the project
3. Quick Use¶
Inference model download¶
This section takes the trained end-to-end model as an example to quickly use the model prediction. First, download the trained end-to-end inference model download address
- In Windows environment, if 'wget' is not installed, the link can be copied to the browser when downloading the model, and decompressed and placed in the corresponding directory
After decompression, there should be the following file structure:
Single image or image set prediction¶
Visualization results¶
The visualized end-to-end results are saved to the ./inference_results folder by default, and the name of the result file is prefixed with 'e2e_res'. Examples of results are as follows:
4. Model Training,Evaluation And Inference¶
This section takes the totaltext dataset as an example to introduce the training, evaluation and testing of the end-to-end model in PaddleOCR.
Data Preparation¶
Download and unzip totaltext dataset to PaddleOCR/train_data/, dataset organization structure is as follow:
total_text.txt: the format of dimension file is as follows，the file name and annotation information are separated by "\t":
The image annotation after json.dumps() encoding is a list containing multiple dictionaries.
The points in the dictionary represent the coordinates (x, y) of the four points of the text box, arranged clockwise from the point at the upper left corner.
transcription represents the text of the current text box. When its content is "###" it means that the text box is invalid and will be skipped during training.
If you want to train PaddleOCR on other datasets, please build the annotation file according to the above format.
Start Training¶
PGNet training is divided into two steps: Step 1: training on the synthetic data to get the pretrain_model, and the accuracy of the model is still low; step 2: loading the pretrain_model and training on the totaltext data set; for fast training, we directly provide the pre training model of step 1download link.
If CPU version installed, please set the parameter use_gpu to false in the configuration.
In the above instruction, use -c to select the training to use the configs/e2e/e2e_r50_vd_pg.yml configuration file.
For a detailed explanation of the configuration file, please refer to config.
You can also use -o to change the training parameters without modifying the yml file. For example, adjust the training learning rate to 0.0001
Load trained model and continue training¶
If you would like to load trained model and continue the training again, you can specify the parameter Global.checkpoints as the model path to be loaded.
Note: The priority of Global.checkpoints is higher than that of Global.pretrain_weights, that is, when two parameters are specified at the same time, the model specified by Global.checkpoints will be loaded first. If the model path specified by Global.checkpoints is wrong, the one specified by Global.pretrain_weights will be loaded.
PaddleOCR calculates three indicators for evaluating performance of OCR end-to-end task: Precision, Recall, and Hmean.
Run the following code to calculate the evaluation indicators. The result will be saved in the test result file specified by save_res_path in the configuration file e2e_r50_vd_pg.yml
When evaluating, set post-processing parameters max_side_len=768. If you use different datasets, different models for training.
The model parameters during training are saved in the Global.save_model_dir directory by default. When evaluating indicators, you need to set Global.checkpoints to point to the saved parameter file.
Model Test¶
Test the end-to-end result on a single image:
Test the end-to-end result on all images in the folder:
Model inference¶
(1).Quadrangle text detection model (ICDAR2015)¶
First, convert the model saved in the PGNet end-to-end training process into an inference model. In the first stage of training based on composite dataset, the model of English data set training is taken as an examplemodel download link, you can use the following command to convert:
For PGNet quadrangle end-to-end model inference, you need to set the parameter --e2e_algorithm="PGNet" and --e2e_pgnet_valid_set="partvgg", run the following command:
The visualized text detection results are saved to the ./inference_results folder by default, and the name of the result file is prefixed with 'e2e_res'. Examples of results are as follows:
(2). Curved text detection model (Total-Text)¶
For the curved text example, we use the same model as the quadrilateral
For PGNet end-to-end curved text detection model inference, you need to set the parameter --e2e_algorithm="PGNet" and --e2e_pgnet_valid_set="totaltext", run the following command:
The visualized text detection results are saved to the ./inference_results folder by default, and the name of the result file is prefixed with 'e2e_res'. Examples of results are as follows: