Writer recognition of Arabic text by generative local features

The generative model of computer vision, along with local features represented by quantized Scale Invariant Feature Transform (SIFT) descriptors, are used to classify writers based on images taken of Arabic text documents. It is the first known application of the method to automated writer recognition. This statistically based approach does not exploit spatial relationships among image features, nor demand explicit segmentation of linguistic units, and does not require supervised training or pre-processing. A performance of 98.0% correct Rank-1 retrieval was achieved for 51 writers, each of whom wrote three cursive samples of the "Rabbit Letter." Although the text of each document in this study was the same, the techniques reported here are designed to be text independent.