Text independent writer identification of Arabic manuscripts and the effects of writers increase

Abstract

This article addresses text-independent writer identification of Arabic manuscripts. Several types of statistical features are extracted from historical Arabic manuscripts. Gradient distribution features for Arabic handwritten text as well as windowed gradient distribution features, contour chain code distribution features, and windowed contour chain code distribution features are extracted. A nearest neighbor (NN) classifier is used with the Euclidean distance measure. Due to the lack of publicly available Arabic manuscript database, this work designed and collected a database of 10,000 Arabic manuscript images handwritten by 200 different historical scholars. Using 8,000 images for training and 2,000 images for testing, the proposed writer identification classifier achieved a top-1, top-5, and top-10 recognition rates of 93.95%, 98.30%, and 99.10%, respectively. The effects of increasing the number of writers on the accuracy results are presented and analyzed.