Local Normalization Towards Global Recognition of Arabic Handwritten Script

In this paper we present a geometrical correction method of handwritten characters. This normalization method is based on Fourier coefficients of a chain-encoded contour. It is carried out in order to improve the rate and the performance of an Arabic handwritten word recognition system. The recognition is based on the global observation of apparent features associated to key-letters of the recognized word, followed by a succession of local observations of secondary features in the word. The global observation does not only try to detect key-word letters, but it also generates a set of characters giving some hints about the identity of the unknown letters present in the word image. The normalization stage leads back the features matching the letter hypothesis in order to increase the chance of recognizing the word presented to the system. Transformed parameters are the angular rotation of the first harmonic phasors and the magnitude of the semimajor axis associated to Fourier coefficients of the truncated approximation to a closed contour. Experimental tests help to evaluate distances between the normalized character and its reference.