✨ 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 (See Release)
You can find the rest codes in TPGSR, TATT and STT.
Current codes may have some bugs, I will fix them soon. Thanks.
📧 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 65.7%/55.1%/38.9%.
evaling easy
[2022-11-19 16:59:48] PSNR 21.74 | SSIM 0.8631
save display images
sr_accuray_iter0: 65.66%
lr_accuray: 37.49%
hr_accuray: 76.41%
best_easy = 65.66%*
evaling medium
[2022-11-19 17:00:09] PSNR 18.21 | SSIM 0.6330
save display images
sr_accuray_iter0: 55.07%
lr_accuray: 21.40%
hr_accuray: 75.05%
best_medium = 55.71%
evaling hard
[2022-11-19 17:00:28] PSNR 19.23 | SSIM 0.7065
save display images
sr_accuray_iter0: 38.87%
lr_accuray: 21.15%
hr_accuray: 64.56%
best_hard = 38.87%*
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!