Application of Natured-Inspired Technique to Odia Handwritten Numeral Recognition

Introduction

Over the last ten decades optical character recognition (OCR) is the most demanding area which comes under the field of pattern recognition, artificial intelligence and machine vision. There are varieties of applications of handwritten character such as postal pin code verification, passport verification, digital library system, document processing, forgery detection in banks and many more. OCR is the translation of handwritten and printed images of character into editable character. The challenging task lies in the recognition of optically processed characters. In off-line recognition method characters are recognized later, after their creation. But the characters are recognized immediately after their creation in an on-line method. OCR can recognize both printed and handwritten characters, but the efficiency of recognition precisely depends on the constraints associated with the characters. More constrained the character is, greater is the efficiency of the OCR system. It is very difficult to develop OCR system for totally unconstrained handwritten characters. Recognition of handwritten characters and numerals is more complex task than the printed character. Many variations are observed in handwritten character due to the style of writing, added noise, missing part, devices and medias used for image acquisition. These may be due to different size, stroke and slant of the image. Therefore there is a requirement to develop robust system which can recognize effectively any handwritten characters. This chapter develops a recognition model for offline handwritten Odia numerals using FLANN as classifier and a bio-inspired technique such as differential evolution (DE) for the weight optimization of the FLANN model. Preprocessing is the first step of OCR followed by feature extraction and classification. Before developing the recognition model preprocessing and feature extraction tasks are performed. The numeral images are preprocessed first. Then feature based recognition method is applied to measure and extract significant features from the numerals image. Then a hybrid system is developed by using FLANN classifier where the weights of FLANN classifier are optimized with DE algorithm. The features are compared to the prototypes developed in training phase. The description which provides the closest match provides the recognition. For feature extraction the recently developed S-transform (M. Hariharan et.al, 2014) is applied which retains the significant features of the pattern. These features being more in dimensions are then reduced by applying principal component analysis (PCA). For the classification task FLANN model is used, the weights of which are adjusted by minimizing the squared error value using the DE algorithm.