6.2 KiB
Algorithm introduction
[TOC]
1. Text Detection Algorithm
PaddleOCR open source text detection algorithms list:
On the ICDAR2015 dataset, the text detection result is as follows:
| Model | Backbone | precision | recall | Hmean | Download link |
|---|---|---|---|---|---|
| EAST | ResNet50_vd | 88.18% | 85.51% | 86.82% | Download link |
| EAST | MobileNetV3 | 81.67% | 79.83% | 80.74% | Download link |
| DB | ResNet50_vd | 83.79% | 80.65% | 82.19% | Download link |
| DB | MobileNetV3 | 75.92% | 73.18% | 74.53% | Download link |
| SAST | ResNet50_vd | 92.18% | 82.96% | 87.33% | Download link |
On Total-Text dataset, the text detection result is as follows:
| Model | Backbone | precision | recall | Hmean | Download link |
|---|---|---|---|---|---|
| SAST | ResNet50_vd | 88.74% | 79.80% | 84.03% | Download link |
Note： Additional data, like icdar2013, icdar2017, COCO-Text, ArT, was added to the model training of SAST. Download English public dataset in organized format used by PaddleOCR from Baidu Drive (download code: 2bpi).
For use of LSVT street view dataset with a total of 3w training data，the related configuration and pre-trained models for text detection task are as follows:
| Model | Backbone | Configuration file | Pre-trained model |
|---|---|---|---|
| ultra-lightweight OCR model | MobileNetV3 | det_mv3_db.yml | Download link |
| General OCR model | ResNet50_vd | det_r50_vd_db.yml | Download link |
- Note: For the training and evaluation of the above DB model, post-processing parameters box_thresh=0.6 and unclip_ratio=1.5 need to be set. If using different datasets and different models for training, these two parameters can be adjusted for better result.
For the training guide and use of PaddleOCR text detection algorithms, please refer to the document Text detection model training/evaluation/prediction
2. Text Recognition Algorithm
PaddleOCR open-source text recognition algorithms list:
Refer to DTRB, the training and evaluation result of these above text recognition (using MJSynth and SynthText for training, evaluate on IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE) is as follow:
| Model | Backbone | Avg Accuracy | Module combination | Download link |
|---|---|---|---|---|
| Rosetta | Resnet34_vd | 80.24% | rec_r34_vd_none_none_ctc | Download link |
| Rosetta | MobileNetV3 | 78.16% | rec_mv3_none_none_ctc | Download link |
| CRNN | Resnet34_vd | 82.20% | rec_r34_vd_none_bilstm_ctc | Download link |
| CRNN | MobileNetV3 | 79.37% | rec_mv3_none_bilstm_ctc | Download link |
| STAR-Net | Resnet34_vd | 83.93% | rec_r34_vd_tps_bilstm_ctc | Download link |
| STAR-Net | MobileNetV3 | 81.56% | rec_mv3_tps_bilstm_ctc | Download link |
| RARE | Resnet34_vd | 84.90% | rec_r34_vd_tps_bilstm_attn | Download link |
| RARE | MobileNetV3 | 83.32% | rec_mv3_tps_bilstm_attn | Download link |
| SRN | Resnet50_vd_fpn | 88.33% | rec_r50fpn_vd_none_srn | Download link |
Note： SRN model uses data expansion method to expand the two training sets mentioned above, and the expanded data can be downloaded from Baidu Drive (download code: y3ry).
The average accuracy of the two-stage training in the original paper is 89.74%, and that of one stage training in paddleocr is 88.33%. Both pre-trained weights can be downloaded here.
We use LSVT dataset and cropout 30w training data from original photos by using position groundtruth and make some calibration needed. In addition, based on the LSVT corpus, 500w synthetic data is generated to train the model. The related configuration and pre-trained models are as follows:
| Model | Backbone | Configuration file | Pre-trained model |
|---|---|---|---|
| ultra-lightweight OCR model | MobileNetV3 | rec_chinese_lite_train.yml | Download link |
| General OCR model | Resnet34_vd | rec_chinese_common_train.yml | Download link |
Please refer to the document for training guide and use of PaddleOCR text recognition algorithms Text recognition model training/evaluation/prediction