Offline Handwriting Arabic Words features Extractions and classifications using Hybrid Transform and Self-Organizing Feature Map

Abstract

In this paper images of Arabic handwritten words have been introduced to sequences of transforms, to obtain the final features of the given image. The hybrid transform has been used which transforms images of Arabic handwritten words into another form such that it is partially invariant to scale and rotation. The transforms used in this research, such as 2-D Fast Fourier Transform, Radon Transform, 1-D Inverse Fast Fourier Transform and 1-D Discrete Multiwavelete Transform have been considered for feature extraction. In the hybrid transform, images of the Arabic handwritten words have been introduced to the sequences of transforms, to obtain the final coefficients matrix of the given image. The obtained transform coefficients can be as affine invariant pattern features. The experiments showed that a small subset of these coefficients is enough for reliable recognition of complex patterns. A Comprehensive Database of Handwritten Arabic Words [1] had been used in the implementation of the hybrid transform. Then Kohonen Self Organizing Feature Map (SOFM) for pattern clustering has been used to cluster the discriminated information vectors extracted from the hybrid transforms matrix’s coefficients. The variations of the transforms used in order to improve generalization, and perform with 89% accuracy on a 7-class lexicon. The technique suggested in this paper can serve as a pre-processing step in computer vision applications.