Arabic handwriting recognition using sequential minimal optimization

Abstract

Due to the variability of writing styles and to other problems related to the nature of Arabic scripts, the recognition of Arabic handwriting is still awaiting accurate results. Segmentation of Arabic handwritten words into graphemes poses a major challenge in Arabic handwriting recognition and is highly error prone. In this paper, we adopt the holistic approach which handles the whole word image without any segmentation step. A set of different statistical features were investigated in this paper, namely, the Invariant Moments (IV), Histogram of Oriented Gradients (HOG) and the Gabor features. The classifier used is the Sequential Minimal Optimization (SMO) algorithm which is an improvement of the Support Vector Machines (SVM). The dataset used is AHDB which consists of 3045 images containing the most commonly used Arabic words written by one hundred different writers. The application of the features used with the SMO algorithm resulted in 91.5928 % correct classification.