Finding the optimum classifier: Classification of segmentable components in offline handwritten Devanagari words

Majority of the approaches towards handwritten Devanagari character segmentation are mostly applied on the word image as a whole. Most of the times, a handwritten word contains some parts that need not require segmentation. If such parts of a word are pre-identified before segmentation, then these can be separated from the word before applying any segmentation algorithm to reduce the entire overhead of the segmentation methodology. In this paper, we propose an approach that finds the number of components present in a word image. Cumulative stretch and shadow based features are extracted from each of these components. Then a range of classifiers are used to see if the component requires further segmentation or not. The parameters of these classifiers are tuned and the results, thus obtained are compared to see which classifier is best suited for such classification. This method is observed to provide an accuracy of 98.63%. We have compared the accuracy of our proposed method with recent works to show the efficacy of our proposed method.