Telugu OCR using Dictionary Learning and Multi-Layer Perceptrons

Abstract

Dictionary Learning (DL) methods have been applied successfully in image processing applications like image de-noising, inpainting, mostly using their representation and reconstruction capabilities. From our experimentation and also from literature, it is seen that the performance of DL approaches on classification tasks is not satisfactory. Especially, the performance is seen to degrade with higher dimensionality and increasing number of classes. We propose a hybrid approach to overcome the classification problem encountered by DL approaches. The novel approach uses the strengths of data abstraction and reconstruction from the DL methods while realising a high classification accuracy through a simple Multi-Layer Perceptron (MLP). In the proposed approach, data abstraction is achieved by the DL method and learned sparse codes are used as inputs for training the MLP. The training is relatively fast as the entire dataset need not be trained. The objective is to minimize the computational requirements for classifying complex datasets like Telugu OCR, without compromising the performance. The method has been tested on University of Hyderabad Telugu Printed Connected Components (UHTelPCC) and Modified National Institute of Standards and Technology database (MNIST) datasets with results comparable to the state-of-the-art methods.