INTEGRATING PART LEARNING WITH STRUCTURE LEARNING FOR HANDWRITTEN NUMERAL RECOGNITION

Abstract

Handwritten Numeral Recognition is the task of correctly identifying handwritten digits. It has an important use-case in digitising old script, image restoration. The challenges while performing handwritten digits recognition arise due to the fact that every person's handwriting can differ. This difference in handwriting can be due to different fonts, slant in letters. Bad quality of text or image from which digits also creates difficulty in recognition. Handwritten Numeral Recognition has been an active area of research for decades but still using part-learning to solve this problem has hardly been explored. Part-learning is a technique in which the original image is divided into parts (called patches), and these patches are learned. Focusing on patches instead of whole images help identify patterns which are sometimes missed. It makes the identification process resistant to background noise. Structure learning paradigm in which we train neural networks by leveraging structured learned from the neural network itself. Structure learned in structure learning can be vectors, graphs or even finite state machines. In this work, we propose a novel approach of integrating part-learning with structure learning for handwritten numeral recognition. Convolutional Neural Network has been used extensively used in the vision-related tasks. We have tested our approach with a multilayer perceptron, convnets and autoencoders. Comparison of the performance of handwritten numeral recognition on MNIST dataset between state-of-the-art techniques and our proposed method indicate the efficacy of our approach.