This software implements the Convolutional Recurrent Neural Network (CRNN) in caffe. Origin software could be found in crnn .You can also find the cnn.pytorch
The initial environment for this project is warpctc-caffe. The network framework is from cnn.pytorch ,it also can be read from crnn
CNN-LSTM-CTC, CNN can get BxC feature vectors,they are as the input of LSTM,LTSM can work the vertors in time_step.From CRNN you can get better way to work as the end-to-end character recognition.At first,I just use the LSTM-CTC( warpctc-caffe ) to work with the license plate recognition.Loss is very slow convergence and it always overfitting.The result don't like the cnn.pytorch ,I run the crnn.pytorch to work with these license plate.Finally,I am determined to write the crnn.caffe.The caffe environment is easier to make a product.
clone the code
git clone [email protected]:yalecyu/crnn.caffe.git
cd crnn.caffe
make
If there some problems,you can confirm your basic the setting of caffe and server.Build this project is the same way like build caffe.
For example,I run a smaill captcha model.You can use like: ./build/examples/cpp_recognition/recognition.bin (your).jpg (your).prototxt (your).caffemodel
./build/examples/cpp_recognition/recognition.bin data/captcha/1.png examples/crnn/deploy.prototxt examples/crnn/model/crnn_captcha.caffemodel
"- - - - - 8 8 - - - - 0 - - - - 3 3 - - - - 5 - - - - - 9 9 - -"
the ans is 80359
You can change some parts of the examples/crnn/deploy.prototxt & examples/cpp_recognition/recognition.cpp & examples/crnn/crnn.prototxt,to suit your project.
python examples/crnn/generate_captcha.py
python examples/crnn/generate_dateset.py
- You can read the image at data/captcha/
./examples/crnn/train.sh
- a cup of caffe
- You can get the more information by reading code
The authors would like to thank the developers of crnn cnn.pytorch warpctc-caffe,without your help ,I can't write the crnn.caffe.When I send the email to you,always get the suggest quickly.
Please let me know if you encounter any issues.