xiongshufeng/DBCAN

To Fuse Semantic and Positional Clues with Cross-Attention for Scene Text Recognition

Python

DBCAN: Dual-Branch Cross-Attention Newtwork for Scene Text Recognition

Paper Link

Requirements

conda create -n DBCAN python=3.8.10

conda install pytorch=1.8.0 cudatoolkit=11.0 torchvision=0.9.0

pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu110/torch1.8.0/index.html

pip install mmdet

git clone git@github.com:Gaoxj2020/DBCAN.git

cd DBCAN && pip install -r requirements.txt

Data Preparing

For Training data and Test data, just follow the instructions in https://mmocr.readthedocs.io/en/latest/datasets/recog.html
Specially,for IC15 and IC13, We use the protocol proposed in [1], we provide the two datasets on
BaiduYun key:u2jh

For Test

We provide code for using our pretrained model to recognize text images.

Our model can be downloaded via Baidu net disk: download_link key: aeo1
Edit the file configs/DBCAN.py Make sure the paths of test datasets are right.
Copy the path of the pretrained model && run bash tools/dist_test.sh configs/DBCAN.py [THE MODEL PATH] [GPU_NUMs]

For Train

Two simple steps to train your own model:

To run the training code, please modify configs/DBCAN.py to your own training data path. Make sure all the paths of Training set and Test set are right.
Run bash tools/dist_train.sh configs/DBCAN.py [OUTPUT PATH] [GPU_NUMs]

Acknowledgement

The code of this project is modified from MMOCR.

[1] Kai Wang, Boris Babenko, and Serge Belongie. End-to-endscene text recognition. In ICCV, pages 1457–1464. IEEE,2011.