This is the code for AAAI2022 paper "Text Gestalt: Stroke-Aware Scene Text Image Super-Resolution". The paper is available at link.

• Publish pre-trained weights and logs
• Make this paper publicly available
• Upload more datasets and stroke-level decomposition rules
• Upload the setups of more recognizers like ASTER, NRTR, etc.

Build up an environment with python3.6, and download corresponding libraries with pip

pip install -r requirement.txt

Download the pre-trained weights and logs at BaiduYunDisk with password: vqg7

Download all resources at BaiduYunDisk with password: mteb

All the resources shoulded be placed under ./dataset/mydata, for example

./dataset/mydata/train1
./dataset/mydata/train2
./dataset/mydata/test
./dataset/mydata/english_decomposition.txt
./dataset/mydata/crnn.pth
./dataset/mydata/pretrain_transformer_stroke_decomposition.pth

Please remember to modify the experiment name. Two text-focused modules are activated whenever --text_focus is used

CUDA_VISIBLE_DEVICES=GPU_NUM python main.py --batch_size=16 --STN --exp_name EXP_NAME --text_focus 

CUDA_VISIBLE_DEVICES=GPU_NUM python main.py --batch_size=16 --STN --exp_name EXP_NAME --text_focus --resume YOUR_MODEL --test --test_data_dir ./dataset/mydata/test

We inherited most of the frameworks from TextZoom and use the pretrained CRNN model from CRNN. Thanks for your contribution!

@JasonBoy1

@meijieru

Will be available soon