The official code of TPS_PP (IJCAI 2023) Paper Link

TPS++, an attention-enhanced TPS transformation that incorporates the attention mechanism to text rectification for the first time. TPS++ builds a more flexible content-aware rectifier, generating a natural text correction that is easier to read by the subsequent recognizer. This code is based on MMOCR 0.4.0 ( Documentation ) with PyTorch 1.6+.

• NRTR + TPS_PP
• CRNN + TPS_PP
• ABINet-LV + TPS_PP

Please refer to Install Guide.

Please see Getting Started for the basic usage of MMOCR 0.4.0.

The specific configuration of the dataset for training and testing can be found here Dataset Document

testing 
├── mixture
│   ├── icdar_2013
│   ├── icdar_2015
│   ├── III5K
│   ├── ct80
│   ├── svt
│   ├── svtp

training
├── mixture
│   ├── Syn90k
│   ├── SynthText

Get the pretrained models from BaiduNetdisk(passwd:cd9r), GoogleDrive. checkpoint model in model/xxx/latest.pth, pre-train model in pre_train/xxx/latest.pth

First, the model needs to be pre-trained using without TPS_PP (pre-train), and then trained end-to-end with a network that incorporates TPS_PP (checkpoint). * denotes the performance of the implemented code. checkpoint model in model/xxx/latest.pth, pre-train model in pre_train/xxx/latest.pth.

Please refer to the training configuration Training Doc

Setp 1 : Download NRTR pre_train/nrtr/latest.pth in mmocr_ijcai/nrtr/latest.pth

#Step 2
PORT=1234 ./tools/dist_train.sh configs/textrecog/nrtr/nrtr_tps++.py ./ckpt/ijcai_nrtr_tps_pp 4 
          --seed=123456 --load-from=mmocr_ijcai/nrtr/nrtr_latest.pth

Please refer to the testing configuration Testing Doc

This code is based on MMOCR

If you find our method useful for your reserach, please cite

@article{zheng2023tps++,
  title={TPS++: Attention-Enhanced Thin-Plate Spline for Scene Text Recognition},
  author={Zheng, Tianlun and Chen, Zhineng and Bai, Jinfeng and Xie, Hongtao and Jiang, Yu-Gang},
  journal={IJCAI},
  year={2023}
}

This project is released under the Apache 2.0 license.