Crnn_Ctpn_Unified

Unified text detection and recognition

First of all, thanks for the author of crnn and ctpn algorithm with opening their codes. Here is crnn original codes, here is ctpn original codes and a tensorflow version can be found here

In my repository, ctpn was built in tensorflow, meanwile crnn was built in pytorch

Environment

Ubuntu16.04(it's more appropriate to build warp-ctc than other)

python3

tensorflow-gpu(version>=1.6)

pytorch(version>=0.3)

warp-ctc click here to download, follow the author to build warp-ctc

cuda 9.0 and cudnn 7.0

Just using pip to establish your dev environment, eg. pip install tensorflow-gpu==1.6, pip install torchvesion(it will install pytorch meanwhile)

Performance

Images and results will be stored in detect_data file.

cd end2endDec_Rec

python3 ./ctpn_crnn_combined/end2endDet_Rec.py

Results

following is image examples

results were saved in detect_data/txt_results

Training data

Here I just share the method that was used to train a crnn model. In my training, ultimately it reached an accuracy of 97.8%. Before training, you should generate a lmdb dataset. Click here for downloading a 3.6 million chinese characters training sets. Or just download the lmdb datasets that were generated by myself.

Here are some examples about chinese characters training data.

if you want to create it by yourself, you can reference dataset_preprocess directory.

Train crnn models

cd train_crnn

python3 crnn_main.py --trainroot lmdb_data_set/xxx --valroot lmdb_data_set/xxx (--cuda)

Here are some software version pictures