CRNN_Chinese_HandWriting_Text_Recognition

This project is modified from https://github.com/Sierkinhane/crnn_chinese_characters_rec

Dependence

• warp_ctc_pytorch
• lmdb

Data Generator

The Training dataset is from, use following commands to download

wget http://www.nlpr.ia.ac.cn/databases/download/feature_data/HWDB1.1trn_gnt.zip
wget http://www.nlpr.ia.ac.cn/databases/download/feature_data/HWDB1.1tst_gnt.zip

After Extract the files, you will find another compressed files in alz format.

In Ubuntu, run unalz HWDB1.1trn_gnt.alz and unalz HWDB1.1tst_gnt.alz

run data_generator/preProcessing.py to extract all 3755 Chinese characters and generate charDict.npy which is the dictionary mapping folder name to Chinese Characters

---

To generate texts in Chinese, Randomly pick chars from dataset

run data_generator/chinese_text.py

data_generator/charDict.npy is the char dictionary

---

Here are some samples.

in test_width.txt

00001.jpg 符疆葛去卑狂擅改汐堂苯谎粥紫鸣

Attention ! To generate lmdb file for crnn training, chinese characters should be stored in UTF-8 format. Try to use NodePad++ to transform ANSI format to UTF-8 format

In alphabets.py, there are two chinese character dictionaries. Switch between each other if you found IndexError

For Training, Firstly, run tolmdb_py3.py to compress all data and labels in lmdb format, then

python crnn_main.py

Results

The Training Set has 200, 000 Chinese text images. The regression is really slow. It toke around 50 hours on RTX2080Ti

Training Loss:

Testing Accuracy:

Example 1

Example 2

Example 3

In TestAll.py you can simply test the whole dataset with Batch_Test_Flag = True

Test the Randomly Generated Dataset, the accuracy is around $$90%$$