Evaluation of Similarity Measures for Recognition of Handwritten Kannada Numerals

Abstract

The automatic classification of patterns is a broad area of research in the machine learning area. The aim of pattern classification is the allocation of a certain input to a specific class in a predefined set of classes. Examples of pattern classification tasks are automatic identification of diseases based on a set of symptoms, optical character recognition, automatic document classification, speech recognition, etc.,. In classification problems, the classification rates depend significantly on similarity measures. Classification depends largely on distance or similarity as neighbors are different depending on similarity measures. Therefore it is important to choose a suitable similarity measure. In this paper an Evaluation of four different similarity measures such as Euclidean, Chebyshev, Manhattan and cosine for recognition of Handwritten Kannada numerals have been done. Here, image fusion technique has been used where extracted features of the several images corresponding to each handwritten numeral are fused to generate patterns, which are stored in 8x8 matrices, irrespective of the size of images. Zonal based feature extraction algorithm is being used to extract the features of Handwritten Kannada numerals. The numerals to be recognized are matched using nearest neighbor classifier with different similarity measures against each pattern and the best match pattern is considered as the recognized numeral. Results show that Euclidean distance measure outperforms other similarity measures in terms of recognition accuracy.