naveen-marthala / Handwritten-word-recognition-OCR----IAM-dataset---CNN-and-BiRNN

handwritten word recognition with IAM dataset using CNN-Bi-LSTM and Bi-GRU implementation.

This project is about recognising handwritten words with CNN and Bi-directional GRU, decoded with CTC.

Dataset:

The IAM Handwriting dataset I have used contains 115,320 isolated and labeled images of words by 657 seperate writers.

IAM words dataset can be downloaded from here. There's also a labelled dataset available for images of lines.

Results:

Test image following the predicted text are shown below:

all



Kings



SGhigraphies



and



SI



SIt



the



show



Gcertain

Yes, the results aren't very promising and only about 59% of the images in test set were identified correctly out of all images of words in the test/unseen set. I presume this is happening because of something improper in gates of GRU.

Although such mistakes in spellings can be corrected using a language model. My colab session had crashed (12.72GB of RAM filling up completely) everytime I tried to import pre-trained language model(I was trying to use 'Google Billion words' dataset). And for this reason, I have uploaded the jupyter notebooks without having corrected the spellings. Yes, I do have plans to fix this in the future using Virtual Machines on cloud.

Training:

Trained on GPU on Google Colab with tensorflow.keras and took around 9 hours to complete.

References and Thanks:

1. Image Pre-processing was partly inspired from: OCR example on keras github repo.
2. Custome CTC Loss function from this article.
3. Network architecture was inspired from following repositories:
   • SimpleHTR,
   • A-CRNN-model-for-Text-Recognition-in-Keras,
   • HTR-for-IAM