A Novel Comparison between the Performances of Several Invariant Moments for Printed Multi-Oriented, Multi-Scaled and Noisy Eastern Arabic Numerals Recognition

In this paper, we present a novel comparison between the robustness against noise of Hu, Legendre, pseudo-Zernike and Krawtchouk invariant moments and even more of invariant analytical Fourier-Mellin transform for multi-oriented, multi-scaled and noisy printed Eastern Arabic numerals recognition. These descriptors are used to extract the features from all numeral images. For this purpose in order to pre-process each one of them, we have used the median filter and the thresholding technique for enhancing its quality, while for recognizing each unknown numeral we have exploited the support vectors machine. Furthermore for carrying out efficiently this comparison, we introduce new concepts which are the threshold and the interval of stability of each invariant descriptor and for each Eastern Arabic numeral. The experiments that we have obtained have provided very satisfactory results.