Introduction
GeneralOCR is open source Optical Character Recognition based on PyTorch. It makes a fidelity and useful tool to implement SOTA models on OCR domain. You can use them to infer and train the model with your customized dataset. The solution architecture of this project is re-implemented from facebook Detectron and openmm-cv.
Installation
Refer to the guideline of gen_ocr installation
Inference
Configuration
Model text detection
Supported Algorithms:
Text Detection
| Algorithm | Paper | Python argument (--det) |
|---|---|---|
| - [x] DBNet (AAAI'2020) | https://arxiv.org/pdf/1911.08947 | DB_r18, DB_r50 |
| - [x] Mask R-CNN (ICCV'2017) | https://arxiv.org/abs/1703.06870 | MaskRCNN_CTW, MaskRCNN_IC15, MaskRCNN_IC17 |
| - [x] PANet (ICCV'2019) | https://arxiv.org/abs/1908.06391 | PANet_CTW, PANet_IC15 |
| - [x] PSENet (CVPR'2019) | https://arxiv.org/abs/1903.12473 | PS_CTW, PS_IC15 |
| - [x] TextSnake (ECCV'2018) | https://arxiv.org/abs/1807.01544 | TextSnake |
| - [x] DRRG (CVPR'2020) | https://arxiv.org/abs/2003.07493 | DRRG |
| - [x] FCENet (CVPR'2021) | https://arxiv.org/abs/2104.10442 | FCE_IC15, FCE_CTW_DCNv2 |
Table 1: Text detection algorithms, papers and arguments configuration in package.
Model text recognition
Text Recognition
| Algorithm | Paper | Python argument (--recog) |
|---|---|---|
| - [x] CRNN (TPAMI'2016) | https://arxiv.org/abs/1507.05717 | CRNN, CRNN_TPS |
| - [x] NRTR (ICDAR'2019) | https://arxiv.org/abs/1806.00926 | NRTR_1/8-1/4, NRTR_1/16-1/8 |
| - [x] RobustScanner (ECCV'2020) | https://arxiv.org/abs/2007.07542 | RobustScanner |
| - [x] SAR (AAAI'2019) | https://arxiv.org/abs/1811.00751 | SAR |
| - [x] SATRN (CVPR'2020 Workshop on Text and Documents in the Deep Learning Era) | https://arxiv.org/abs/1910.04396 | SATRN, SATRN_sm |
| - [x] SegOCR (Manuscript'2021) | - | SEG |
Table 2: Text recognition algorithms, papers and arguments configuration in package.
Inference
# Activate your conda environment
conda activate gen_ocr
python general_ocr/utils/ocr.py demo/demo_text_ocr_2.jpg --print-result --imshow --det TextSnake --recog SEG
--det and --recog argument values are supplied in table 1 and table 2.
The result as below:
Training
Training with toy dataset
We prepare toy datasets for you to train on /tests/data folder in which you can do your experiment before training with the official datasets.
python tools/train.py configs/textrecog/robust_scanner/seg_r31_1by16_fpnocr_toy_dataset.py --work-dir seg
To change text recognition algorithm into sag:
python tools/train.py configs/textrecog/sar/sar_r31_parallel_decoder_toy_dataset.py --work-dir sar
Training with Academic dataset
When you train Academic dataset, you need to setup dataset directory as this guideline. The main point you should forecus is that your model point to the right dataset directory. Assume that you want to train model TextSnake on CTW1500 dataset, thus your config file of that model in configs/textdet/textsnake/textsnake_r50_fpn_unet_1200e_ctw1500.py should be as below:
dataset_type = 'IcdarDataset'
data_root = 'data/ctw1500/'
data = dict(
samples_per_gpu=4,
workers_per_gpu=4,
val_dataloader=dict(samples_per_gpu=1),
test_dataloader=dict(samples_per_gpu=1),
train=dict(
type=dataset_type,
ann_file=f'{data_root}/instances_training.json',
img_prefix=f'{data_root}/imgs',
pipeline=train_pipeline),
val=dict(
type=dataset_type,
ann_file=f'{data_root}/instances_test.json',
img_prefix=f'{data_root}/imgs',
pipeline=test_pipeline),
test=dict(
type=dataset_type,
ann_file=f'{data_root}/instances_test.json',
img_prefix=f'{data_root}/imgs',
pipeline=test_pipeline))
Your data_root folder data/ctw1500/ have to be right. Afterward, train your model:
python tools/train.py configs/textdet/textsnake/textsnake_r50_fpn_unet_1200e_ctw1500.py --work-dir textsnake
To study other configuration parameters on training.
Testing
Now you completed training of TextSnake and get the checkpoint textsnake/lastest.pth. You should evaluate peformance on test set using hmean-iou metric:
python tools/test.py configs/textdet/textsnake/textsnake_r50_fpn_unet_1200e_ctw1500.py textsnake/latest.pth --eval hmean-iou
Citation
If you find this project is useful in your reasearch, kindly consider cite:
@article{genearal_ocr,
title={GeneralOCR: A Comprehensive package for OCR models},
author={khanhphamdinh},
email= {[email protected]},
year={2021}
}
269 Jan 2, 2023
90 Jan 6, 2023
52 Nov 9, 2022
181 Dec 27, 2022
165 Jan 3, 2023
153 Dec 14, 2022
1 Oct 29, 2021
166 Jan 1, 2023
0 Feb 6, 2022
539 Dec 28, 2022
4 May 28, 2022
290 Dec 25, 2022
1 Jul 14, 2022
74 Sep 22, 2021
136 Dec 31, 2022
6 Jun 08, 2022
7 Sep 20, 2022
1 Nov 19, 2021
26 Dec 05, 2022
4 Dec 03, 2021
64 Jan 05, 2023
10 Nov 09, 2022
102 Dec 06, 2022
51 Dec 11, 2022
5 Jan 05, 2023
2 Jun 05, 2022
571 Dec 28, 2022
378 Dec 30, 2022
2 Oct 09, 2022