Studies on some Soft Computing Techniques: A Case Study for Constrained Handwritten Devnagari Characters and Numerals
Sandhya Arora · 2013
Classifier plays a crucial role in handwritten text recognition system. The major challenge lies in taking the advantage of multiple classifiers by exploiting the strengths of a classifier and by suppressing its weakness by other classifiers. In this paper, some soft computing techniques like Multilayer perceptrons (MLPs), support vector machines (SVM) and method of minimum edit distance has been considered for classification of handwritten isolated Devnagari characters. After preprocessing the character image, eight types of features are extracted. SVM, MLP and its decision combination techniques are explored on eight types of features. Two approaches of two stage classification have also been applied. In first approach, preliminary grouping of characters has been done using structural properties of Devnagari characters. In second stage, MLP is designed for each group of characters using directional chain code features. In second approach, the character patterns are divided into two sets, certainty set and confusing character set. This division is based on a relative difference measure. After separating characters in two sets, a MLP based classifier is used in the first stage to classify characters of certainty set. And for classifying characters of confusing character set, method of minimum edit distance is applied in the second stage, on detected corners of the sample character using a modified form of Harris corner detector. A study of MLP and SVM has also been done for Handwritten Devnagari Numerals. The accuracy of 90.74% and 95.18% is achieved for characters and numerals respectively.