HANDWRITTEN NUMERAL RECOGNITION USING NEURAL NETWORKS

Abstract

Machine learning algorithms including Support Vector Machines (SVM’s), Multilayer Perceptrons and Neural Networks have been successful in vision tasks earlier, but it has been seen that there is stagnation in the error rate or accuracy of these algorithms due to reasons including poor generalization, local minima and weight change. The challenge to improve further still remains. Deep learning has posed several interesting possibilities and state of the art results have been achieved in vision tasks including image labelling, hand written digit recognition etc. Deep learning methods aim at learning feature hierarchies with features from higher levels of the hierarchy formed by the composition of lower level features. Deep architectures attempt to learn hierarchical structures and seem promising in learning simple concepts first and then successfully building up more complex concepts by composing the simpler ones together Many existing learning methodologies have been used in recognition of hindi numerals but none of the deep learning approaches have been tried much in this area as of now. By applying deep learning methods, we are free of hand-crafted low-level features and can automatically learn mid-level and higher-level features from a large amount of unlabelled raw samples beyond types and domains of handwriting recognition also. Through this work we look at various deep learning methodologies and specifically methodologies which help in pre-training of the multiclass classifiers thus avoiding the need of hand crafting features for better results. The idea of deep learning has been implemented using different kinds of neural networks though the concept is not restricted to only neural networks. State-of-the-art results have been achieved on classification experiments performed on hindi numeral images involving techniques of pre-training a deep multiclass classifier basically Autoencoders.