Learning Centric Feature Extraction and Classification Models for OCR

Abstract

Character recognition is the process of enabling computers to classify the characters from their image presentation. The practical application involves bankcheck reading, book reading for the blind, postal address reading, and many more. With emerging trends in technology, mobile devices have been equipped with high definition cameras and powerful processors. This pattern enlarges the scope of mobile applications to read business cards, street signboards, medical prescriptions, and local language translation for travellers, etc. The essential step in all these applications is the recognition of characters. Traditionally the application uses handcrafted features but in recent years, the nonhandcrafted feature extraction methods have gained increasing popularity for solving pattern classification tasks due to their inherent ability to extract robust features and handle outliers. This dissertation focuses on the design of nonhandcrafted learning models for character recognition application. The research primarily involves the proposition of hyperparameterless learning models that can be used for the image classification task. The dissertation proposed three newly developed learning methodology for character classification from raw images. The first contribution overcomes limitations of classification over Single Layer Feed Forward Network. The network is a widely validated model for the classification task. The limitations of such model are due to the need for hyperparameter tuning. Extreme Learning Machine was developed as a hyperparameterless to overcome this limitation. However, the random input weight in Extreme Learning Machine makes it suffer from the illposed problem. Hence, a hyperparameterless algorithm namely BackwardForward Extreme Learning Machine (BFELM) is developed that learns the input and output weights in one backward and one forward pass respectively. The second contribution extended the BFELM framework to learn the weights of the convolutional neural network. The newly developed model is named Convolutional Network with Backward Forward Extreme Learning Machine (CNBFELM). Indepth analysis of the proposed model over the various publicly available dataset, prove its efficiency in hyperparameterless nonhandcrafted learning. The third contribution involves the development of feature learning by optimization. First, feature learning is modeled as an optimization problem that does not depend on classification error or accuracy. It is worth mentioning that the optimization model can generalize even with a smaller number of the training sample. The final contribution is about designing the convolutional neural network model for handwritten word recognition in Indic language. The CMATERdb2.1.2 dataset is used to study the proposed model in depth.