UPOCR: Towards Unified Pixel-Level OCR Interface

🌟 Highlight

The official implementation of UPOCR: Towards Unified Pixel-Level OCR Interface (ICML 2024). The UPOCR represents a first-of-its-kind simple-yet-effective generalist model for unified pixel-level OCR interface. Through the unification of paradigms, architectures, and training strategies, UPOCR simultaneously excels in diverse pixel-level OCR tasks using a single unified model. Below is the framework of UPOCR.

⚒️ Environment

We recommend using Anaconda to manage environments. Run the following commands to install dependencies.

conda create -n upocr python=3.9 -y
conda activate upocr
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu116
git clone https://github.com/shannanyinxiang/UPOCR.git
cd UPOCR
pip install -r requirements.txt

📊 Datasets

• Download the SCUT-EnsText [repo], TextSeg [repo], and Tampered-IC13 [repo] datasets.
• Preprocess the SCUT-EnsText dataset following [link].
• Arrange the datasets according to the file structure below.

data
├─TamperedTextDetection
│  └─Tampered-IC13
│     ├─test_gt
│     ├─test_img
│     ├─train_gt  
│     └─train_img
├─TextRemoval
│  └─SCUT-EnsText
│     ├─train
│     │  ├─image
│     │  ├─label
│     │  └─mask
│     └─test
│        ├─image
│        ├─label
│        └─mask
└─TextSegmentation
   └─TextSeg
      ├─image
      ├─semantic_label
      └─split.json

📺 Inference

• Download the UPOCR weights at [link].
• Run the following command to perform model inference on the TextSeg dataset.

dataset=textseg #  or scut-enstext or tampered-ic13 
output_dir=./output/upocr-infer/

mkdir ${output_dir}

CUDA_VISIBLE_DEVICES=0 \
torchrun \
        --master_port=3140 \
        --nproc_per_node=1 \
        main.py \
        --output_dir ${output_dir} \
        --data_cfg_paths data_configs/train/scut-enstext.yaml data_configs/train/tampered-ic13.yaml data_configs/train/textseg.yaml \
        --eval true \
        --resume pretrained/upocr.pth \
        --eval_data_cfg_path data_configs/eval/${dataset}.yaml \
        --visualize true \
        --textseg_conf_thres 0.4 # Tune this argument for optimal text segmentation performance.

Change the dataset variable to scut-enstext or tampered-ic13 to run inference on the SCUT-EnsText or Tampered-IC13 datasets, respectively.

• For the text removal task, run the following command to calculate image-eval metrics. For the other two tasks, the metrics will be automatically calculated at the above step.

python -u eval/text_removal/evaluation.py \
    --gt_path data/TextErase/SCUT-ENS/test/label/ \
    --target_path output/upocr-infer/SCUT-EnsText

python -m pytorch_fid \
    data/TextErase/SCUT-ENS/test/label/ \
    output/upocr-infer/SCUT-EnsText \
    --device cuda:0

🏋️ Training

• Download the pre-training weights for UPOCR at [link].
• Run the following command for model training.

output_dir=./output/upocr-train/
log_path=${output_dir}log_train.txt

mkdir 'output'
mkdir ${output_dir}

CUDA_VISIBLE_DEVICES=0,1 \
torchrun \
        --master_port=3140 \
        --nproc_per_node=2 \
        main.py \
        --output_dir ${output_dir} \
        --data_cfg_paths data_configs/train/scut-enstext.yaml data_configs/train/tampered-ic13.yaml data_configs/train/textseg.yaml \
        --pretrained_model pretrained/pretraining_weights.pth \
        --amp true | tee -a ${log_path}

✅ Citation

@inproceedings{peng2024upocr,
  title={{UPOCR}: Towards Unified Pixel-Level {OCR} Interface},
  author={Peng, Dezhi and Yang, Zhenhua and Zhang, Jiaxin and Liu, Chongyu and Shi, Yongxin and Ding, Kai and Guo, Fengjun and Jin, Lianwen},
  booktitle={International Conference on Machine Learning},
  year={2024},
}

📇 Copyright

This repository can only be used for non-commercial research purpose.

For commercial use, please contact Prof. Lianwen Jin (eelwjin@scut.edu.cn).

Copyright 2024, Deep Learning and Vision Computing Lab, South China University of Technology.

✨ Star Rising