Arabic/Farsi Handwritten Digit Recognition using Histogram of Oriented Gradient and Chain Code Histogram

Seyyed Khorashadizadeh and Ali Latif

Abstract: The aim of this paper is to propose a novel technique for Arabic/Farsi handwritten digit recognition. We constructed an invariant and efficient feature set by combination of four directional Chain Code Histogram (CCH) and Histogram of Oriented Gradient (HOG). To achieve higher recognition rate, we extracted local features at two levels with grids 2×2, 1×1 and it causes a partial overlapping of zones. Our proposed feature set has 164 dimensions. For classification phase, Support Vector Machine (SVM) with radial basis function kernel was used. The system was evaluated on HODA handwritten digit dataset which consist of 60000 and 20000 training and test samples, respectively. The experimental results represent 99.31% classification rate. Further, 5-fold cross validation was applied on whole 80000 samples and 99.58% accuracy was obtained.

Keywords: Arabic/farsi handwritten digit recognition, CCH, HOG, SVM.

Received November 30, 2014; accepted February 4, 2015