✨ C3-STISR ✨

Official Code for 'C3-STISR: Scene Text Image Super-resolution with Triple Clues'

IJCAI 2022 Accepted Paper

The code will be gradually open source!

Due to company policy, I can't provide the code. To guarantee reproducibility, I am working on reimplement C3-STISR. The final result will be released around September. Thank you all for your support.

Since a lot of my works (including reviewing papers) are written in one framework, I decide not to release the complete code. Instead, following items are released for reproducibility:

• model scrits

• training scripts

• log

• final pth (accessed via email request)

You can find the rest codes in TPGSR, TATT and STT.

📧 Contact 📧

Feel free to contact me if you have any problems! zhaomy20@fudan.edu.cn

😳 Difference between the paper 😳

• Visual clue is removed since it does not help improving recognition performance.

• Linguistical clue is trained with CTC loss.

• Performance of CRNN is boosted from 65.2%/53.6%/39.8% to 65.6%/53.4%/39.9%.

evaling easy
[2022-09-05 14:25:13]	PSNR 22.13 | SSIM 0.8532	
save display images
sr_accuray_iter0: 65.60%
lr_accuray: 37.49%
hr_accuray: 76.41%
best_easy = 65.60%*
evaling medium
[2022-09-05 14:25:29]	PSNR 18.98 | SSIM 0.6449	
save display images
sr_accuray_iter0: 53.44%
lr_accuray: 21.40%
hr_accuray: 75.05%
best_medium = 54.43%
evaling hard
[2022-09-05 14:25:45]	PSNR 19.75 | SSIM 0.7125	
save display images
sr_accuray_iter0: 39.91%
lr_accuray: 21.15%
hr_accuray: 64.56%
best_hard = 39.91%*
saving best model

🔥 Training 🔥

python3 main.py --arch="c3stisr" --test_model="CRNN" --batch_size=48 --STN  --sr_share --gradient  --use_distill --stu_iter=1 --vis_dir='C3-STISR-Final' --mask

💫 Testing 💫

python3 main.py --arch="c3stisr" --test_model="CRNN" --batch_size=48 --STN  --sr_share --gradient  --use_distill --stu_iter=1 --vis_dir='C3-STISR-Final' --mask --go_test --resume='***'

👊 Performance 👊

😆 Citation 😆

If you find this project is useful for your research, please cite:

@inproceedings{zhao2022c3,
  title={C3-STISR: Scene Text Image Super-resolution with Triple Clues},
  author={Zhao, Minyi and Wang, Miao and Bai, Fan and Li, Bingjia and Wang, Jie and Zhou, Shuigeng},
  booktitle={IJCAI},
  pages={1707--1713},
  year={2022}
}

😉 Related Works 😉

· Text Gestalt: Stroke-Aware Scene Text Image Super-Resolution [Paper] [Code]

· A Text Attention Network for Spatial Deformation Robust Scene Text Image Super-resolution [Paper] [Code]

· Scene Text Telescope: Text-Focused Scene Image Super-Resolution [Paper] [Code]

· Text Prior Guided Scene Text Image Super-resolution [Paper] [Code]

👍 Special Thanks 👍

Additionally thank JingyeChen for his help!

Our framework is based on TPGSR, STT, and TG.