【个人开源】2D Attentional Irregular Scene Text Recognizer代码复现

项目地址: https://github.com/chenjun2hao/Bert_OCR.pytorch

Unofficial PyTorch implementation of the paper, which transforms the irregular text with 2D layout to character sequence directly via 2D attentional scheme. They utilize a relation attention module to capture the dependencies of feature maps
and a parallel attention module to decode all characters in
parallel.

At present, the accuracy of the paper cannot be achieved. And i borrowed code from deep-text-recognition-benchmark

model

在这里插入图片描述

result
Test on ICDAR2019 with only 51.15%, will continue to improve.

在这里插入图片描述

Feature

1. Output image string once not like the seqtoseq model

Requirements

Pytorch >= 1.1.0

Test

1. download the pretrained model Baidu password: kdah.

2. test on images which in demo_image folder

python demo.py --image_folder demo_image --saved_model <model_path/best_accuracy.pth>

3. some examples

demo images	Bert_OCR
在这里插入图片描述	available
在这里插入图片描述	shakesshack
在这里插入图片描述	london
在这里插入图片描述	greenstead
在这里插入图片描述	toast
在这里插入图片描述	merry
在这里插入图片描述	underground
在这里插入图片描述	ronaldo
在这里插入图片描述	bally
在这里插入图片描述	university

4. result on benchmark data sets

total_accuracy: 78.423

Train

1. I prepared a small dataset for train.The image and labels are in ./dataset/BAIDU.

python train.py --root ./dataset/BAIDU/images/ --train_csv ./dataset/BAIDU/small_train.txt --val_csv ./dataset/BAIDU/small_train.txt

Reference

1. deep-text-recognition-benchmark
2. 2D Attentional Irregular Scene Text Recognizer