A Novel Approach to Recognize the off-line Handwritten Numerals using MLP and SVM Classifiers

This paper presents a new approach to off-line handwritten numeral recognition. Recognition of handwritten numerals has been one of the most challenging task in pattern recognition. Recognition of handwritten numerals poses serious problems because of high variability in numeral shapes written by individuals. This paper concerns with offline handwritten numeral recognition based on MLP and SVM classifiers. The performance of character recognition system depends heavily on what kind of features extraction techniques are being used. In this work, we have collected 1200 samples of isolated numerals, contributed by 24 writers and each one has written it five times. In this paper we propose a modified Hough transformation technique and four view projection Profiles technique to extract the feature of numerals. Using the modified Hough transformation technique, we have achieved maximum accuracy of about 93.12% and 72.5% by SVM and MLP classifiers and with four view projection profiles technique; we obtained the 96.04% and 98.73% recognition accuracy by SVM and MLP classifiers.