handwritten-sequence-tensorflow

handwritten-sequence-tensorflow copied to clipboard

hello, @johnsmithm ,Good job.

I used to write a project based on CNN+BILSTM+CTC for IAM words, but if failed to converge . So I hope to know this project's performance.

And can you share your data process code for this project, or the traning data file ?

hi, i have used small images with 1-3 words with small letters. width-90 and height - 36 pixels. the best result was 12% for label rate error using overlaping windows of width 30 pixels and step 3. Multidimentional lstm also gives around 15% lre but my implementation is slow, but with nlstm from contrib is faster than overlaping windows model. Now i am working on another repository. Next day i will upload the process files, but i have a small example in the utils.py.

Thanks for your quickly reply, @johnsmithm

Which IAM dataset {IAM-onDB, IAM-DB} is you actually use in this project? I use IAM-DB，it already cut the image which just contain one word in it.

You said you "used small images with 1-3 words with small letters ", does it means you filter all other Captial letters and other symbols.