wjtuz/rfn

Industria Scene Text Detection

Text Detection for Industrial Scene with Refined Feature-attentive Network

Environments

• Ubuntu 16.04
• Cuda 10
• python >=3.5
• pytorch 1.0
• Other packages like cv2, Polygon3, tensorboardX, Scipy.

Highlights

• Training and evaluation checked: Testing in MPSC test set with training data in {SynthMPSC, MPSC}. Other scene text datasets are test with pre-training data in SynthText.
• Dataset link:
  • MPSC&SynthMPSC (Ours) dataset will be released here soon.
  • Synthtext
  • MSRA-TD500
  • ICDAR2013
  • ICDAR2017-MLT
  • USTB-SV1K

Installation

Check INSTALL.md for installation instructions.

Configuring your dataset

• Update datset root path in$RFN_ROOT/train.py.
• Process dataset can be set in $RFN_ROOT/tools/datagen.py.
• Modify test path in $RFN_ROOT/multi_image_test_ocr.py.
• Modify some settings in $RFN_ROOT/tools/encoder.py, including anchor_areas, aspect_ratios.

# refer to /data_process/Compute aspect_ratios and area_ratios.py
self.anchor_areas = [16*16., 32*32., 64*64., 128*128., 256*256, 512*512.]
self.aspect_ratios = [1., 2., 3., 5., 1./2., 1./3., 1./5.,7.]

Training

# create your data cache directory
cd RFN_ROOT
# Download pretrained ResNet50 model(https://data.lip6.fr/cadene/pretrainedmodels/se_resnet50-ce0d4300.pth)
# Init RFN with pretrained ResNet50 model
python ./tools/get_state_dict.py
python train.py --config_file=./configs/R_50_C4_1x_train.yaml

• The training size is set to a multiple of 128.
• Multi-GPU phase is not testing yet, be careful to use GPU more than 1.

Test and eval

• Our provide script: $RFN_ROOT/multi_image_test_ocr.py and $RFN_ROOT/test/
• Modify path settings and choose the dataset you want to evaluate on.
• option parameters: save_img, show_mask

### test each image
python test.py --dataset=MPSC --config_file=./configs/R_50_C4_1x_train.yaml --test --save_img

### eval result
python test.py --dataset=MPSC --eval

Pretrained Weights for Training and Testing

• Here we provide some pretained weights for testing in baidu drive:

  Pretrain SynthMPSC : https://pan.baidu.com/s/1BI2T4ncowKu908dcd9tT7g (0ke0)
  Pretrain SynthText : https://pan.baidu.com/s/16q7jB2vLvW49fcs9nfXxRA (waki)

More Results

• Model | Dataset | Precision | Recall | F-Measure | MODEL link | Extraction code
• RFN | MPSC | 89.30 | 83.33 | 86.21 | model | 6u6y
• RFN* | MPSC | 89.82 | 84.45 | 87.05 | model | xrni

Training logs and results for comparative experimental

  train_log_url : https://pan.baidu.com/s/1364azjk0hdy8Aeekk_tRQg 
  Extraction code : bd9z

• Model | Dataset | Precision | Recall | F-Measure
• MASKRCNN | MPSC | 85.28 | 79.25 | 82.15
• DB | MPSC | 87.77 | 78.73 | 83.00
• PAN | MPSC | 87.07 | 81.60 | 84.24
• PSENET | MPSC | 85.42 | 78.40 | 81.76
• ContourNet | MPSC | 87.79 | 81.02 | 84.27
• RRPN++ | MPSC | 86.73 | 83.90 | 85.30
• FCENET | MPSC | 87.13 | 81.63 | 84.29

Visualizations of MPSC dataset

Visualizations of MSRA-TD500, USTB-SV1K, ICDAR2013, ICDAR2017-MLT dataset

Citation

@article{guan2021industrial,
  title={Industrial Scene Text Detection with Refined Feature-attentive Network},
  author={Guan, Tongkun and Gu, Chaochen and Lu, Changsheng and Tu, Jingzheng and Feng, Qi and Wu, Kaijie and Guan, Xinping},
  journal={arXiv preprint arXiv:2110.12663},
  year={2021}
}