Abstract
This paper presents a novel approach for off-line text-independent Arabic writer identification. The approach operates in four steps: 1) handwritten text is segmented into strokes after an image thinning step; 2)length, height/width ratio and curvature stroke features are extracted; 3) five feature vectors are computed: stroke length/ratio probability distribution function (PDF), stroke length/ratio horizontal and vertical cross-correlation, stroke length/curvature PDF, stroke length/curvature horizontal and vertical cross-correlation, and stroke length/curvature and length/ratio cross-correlation; 4) classification is carried out using different metrics and the Borda count ranking algorithm. A first experimental evaluation performed on 40 writers from the IFN/ENIT database produced a promising identification rate of 92.5% for Top1 and 100% for Top5.