How to train gap-class models?

[breta01]

How to train gap-class models?

[Breta01]

Hi,

I will maybe need a little bit clarification on your question.

Right now there are different approaches to recognition of handwritten text. One of them first separates a word into individual characters using Gap-Classifier and than recognize these separated characters using Char-Classifier. Therefore, Gap-Classifier and Char-Classifier are two different models. If you already have Char-Classifier which can recognize individual characters you need to train Gap-Classifier.

Gap-Classifier takes slices from an image of a word and classifies them on either letters or gaps. Right now I have two different Gap-Classifiers: GapClassifier.ipynb and GapClassifier-BiRNN.ipynb. The first one uses only a convolutional neural network and the second uses it with a combination of a recurrent neural network.

The model you should train depends mainly on your dataset. The GapClassifier.ipynb takes as an input images separated into two classes: the images of letters are in 0 class and the images of gaps between letters have class 1. On the other hand, the GapClassifier-BiRNN.ipynb takes as an input images of words with an array of numbers corresponding to positions of gaps in the image.

Can you give me more details about your dataset, so I can give you more details.

[PR-Iyyer]

Thank you so much @Breta01 .

I wanted the model to predict both lowercase and uppercase characters from even cursive written text. So I made handwritten character dataset (characters alone with corresponding labels) and trained it using CharClassifier.ipynb and obtained a model. This is the model i used for testing a word. I found that the borders drawn for separating characters on a word is wrong. Suppose a word 'Mind' is given, the border is drawn in the middle of character M. Therefore prediction also fails.

I am attaching a snapshot.

SO, can you please guide me with this.

[PR-Iyyer]

SO, can you please guide me with this.

[Breta01]

I understand the problem that we are facing here, but there is a limited number of things I can do about it.

The current GapClassifier which detects the borders between letters was trained on my handwriting and it fails on a handwriting of others. Therefore, if you want to improve the gap detection, you have to extend the current dataset and train new GapClassifier.

There is still a lot of work on this project. One of the most important parts is a large dataset, so I would really appreciate if you were willing to share your dataset.

Hope this explains the issue and thank you for your interest.

[PR-Iyyer]

Thanks @Breta01 . This clarifies . Dataset will upload soon.