How to create a training set for CTC Classifier about handwriting-ocr HOT 3 OPEN

The most important are the word images. You only need the text files if you want to train Seq2seqX model or GapClassifier-BiRNN. Other word classifiers work without the text files, so you can train the CTC model just with the images.

If you really want to use the text files, you can create them manually using WordClassDM.py script. The text files contain an array of positions (x coordinates) of vertical lines which separate individual letters in the corresponding words.

from handwriting-ocr.

Is it important to make every image name unique by adding an random timestamp as I have resized and converted into grayscale. I mean would it create any problem if I run the CTC classifier for it.

Also since I am trying to do it for capital handwritten names often seperated by gaps so should I necessarily use gap classifier to increase accuracy?

from handwriting-ocr.

First, I use the timestamp just to differentiate between images of same word. As long as the file name has the structure <word>_<uniqu_toke>.jpg it should work.

In case you are working only with capital letters it may be worth trying the approach of using gap classifier for separating letters and then recognizing them using char classifier. It mostly depends on your dataset. But if you can separate the letters well there is large letter dataset available online.

from handwriting-ocr.