Improvement of the performance of optical handwritten digit recognition by incorporating cross-domain autoencoder-based image to image translation technique

摘要

The paper represents a novel approach that improves the performance of optical handwritten digit recognition incorporating image to image translation technique in a way - that it translates the input image of handwritten domain to some model domain so that the performance of any specified classier improves. In this paper, we have done image to image translation using two pretrained autoencoders, one of them is an autoencoder of a handwritten domain and other is an autoencoder of a model domain. We have bonded together the encoder of handwritten domain and the decoder of model domain and a neural network for feature translation in between in order to train the whole neural network for the translation of an image from handwritten domain to the image of a model domain. Also, an analysis has been shown regarding how well our method improves the performance of optical handwritten digit recognition for both Bengali and English digits.