Moment based invariant feature extraction techniques for bilingual character recognition

Feature extraction is an important phase in optical character recognition (OCR). Moment based features are very effective in describing shape of characters. In this paper the efficiency of these features for Bilingual Character Recognition (Gurmukhi and Roman) is studied. The detailed analysis for minimizing within-class variability and maximizing between-class variability is studied and it is observed that only a few moments provide this capability. It is observed that moment based features can become very effective if certain operations such as normalization of character size and geometric operations are performed correctly using floating point arithmetic. Based on the analysis of reconstructed images with Zernike moments, pseudo Zernike moments and orthogonal Fourier - Mellin moments, using the first 12, 6 and 7 order of the moments, respectively, it is recommended to compose the feature vectors in order to achieve image recognition results. Pseudo Zernike moments give better results among all types of features Although higher order moments carry more fine details of an image, but they are also more susceptible to noise.