Harald Scheidl

nc: number of classes (characters) to recognize (a, b, c, ...) nl: we need one more pseudo character ("blank") for the CTC layer, therefore nc+1 nt: length of output sequence...

CRNN uses best path decoding: https://github.com/bgshih/crnn/blob/master/src/cpp/ctc.cpp#L152

best path decoding < beam search < beam search with (good) language model

I don't think there is an easy way to "copy" this project to TF. But TF has all components needed out of the box (CNN, LSTM, CTC) which you can...

take a training image (not larger than 100x32px), copy it to a random position in the input image (100x32px), fill the parts of the input image not covered with (white)...

the model does not care what the meaning of a class is, therefore each character is just represented by a number. To input a text string into the model, the...

had the same problem. The reason for this was the wrong version of thpp as far as I remember. Just be sure to use the correct versions for all libs,...

we had our GPU running with 14.04, therefore we had to checkout the non-head versions, see install_all.sh: ``` ... else git clone -b v0.35.0 --depth 1 https://github.com/facebook/folly git clone -b...

the neural network is in fact only interested in labels representing chars (0, 1, 2, ...). Why not simply map each of your arabic chars (UTF8, containing 1 or multiple...

no, sorry. If you can't find a solution for LUA, you could consider switching to a Python/TensorFlow implementation of CRNN (just search for it on github, there should be multiple...