Majlesi Journal of Electrical Engineering Vol. 10, No. 3, September 2016 19 Recognition of Handwritten Persian Two-digit Numerals Using a Novel Hybrid SVM/HMM Algorithm Vahid. Moradi 1 , Farbod. Razzazi 2 , Alireza. Behrad 3 1- Electrical and Computer Engineering Department, Science and Research Branch, Islamic Azad University, Tehran, Iran. Email: v_mo83@yahoo.com 2- Electrical and Computer Engineering Department, Science and Research Branch, Islamic Azad University, Tehran, Iran. razzazi@srbiau.ac.ir 3-Electrical and Electronic Engineering Department, Shahed University, Tehran, Iran. behrad@shahed.ac.ir Received : Oct. 2015 Revised : March 2016 Accepted : June 2016 ABSTRACT: There is a vast range of potential applications for recognition of handwritten Persian / Arabic digits (e.g. banking transactions, business registration forms and postal code recognition to name a few). In this paper, a new method is presented for automatic recognition of joint two-digit Persian numerals. The proposed method is composed of a combinational structure of Support Vector Machines (SVM) and a Hidden Markov Models (HMM). In this approach, we used SVM and HMM for classification and segmentation goals respectively. Due to the higher performance of SVM in classification with respect to HMM, the main core of recognition is a SVM classifier. In contrast, we used HMM to detect the location of the boundary for two-digit numerals. To evaluate the method, we employed a selection of HADAF Persian isolated characters corpus. We employed a 4 scale Gabor filter bank (24, 12, 6 and 3 scales) in 6 directions (0, 30, 60, 90, 120, 150 degrees) for feature extraction. The results showed the digit recognition rate of about 98.75 percent for the proposed algorithm on Persian two-digit numerals, while the recognition rates were 98.58 and 95.93 for separate SVM and HMM engines on isolated characters respectively. KEYWORDS: Persian handwritten numeral recognition, SVM/HMM combining classifier. 1. INTRODUCTION Nowadays, the recognition of handwritten digits has found wide applications. They are applicable to postal zip code reading, handwritten form processing, data entry applications, electronic payment cheques, office automation systems and a large volume of business transactions. The recognition of handwritten Persian isolated numerals, due to high structural similarity of different digits and diversity of handwriting styles faces many challenges. There are valuable glancing results for Persian/Arabic isolated characters. Mahmoud et al address the problem of recognition of Persian/Arabic numerals using SVM and HMM engines separately, the achieved average recognition rates are 99.4%, 97.99% and 94.35% using SVM, HMM and NM classifiers, respectively [1]. Ahranjany et al. used a fusing of two convolutional neural networks (CNN) trained by gradient descent training algorithm for recognizing the handwritten Farsi/Arabic digits. The technique is inspired by the human visual system. This study reports a digit recognition rate of 99.17%. In addition, the recognition rate increases to 99.98% after rejection of ten percent of “hard to recognize” samples [2]. Ching et al used a combining classifier consisting of a CNN and a support vector machine (SVM) for recognition of English digits [3]. In this approach, CNN is employed for feature extraction stage and SVM is used as a classifier for recognition. This algorithm results in an average recognition rate of 99.81%. Yang Zhang et al used a hybrid SVM/HMM for handwritten chemical symbols recognition [4]. In the first stage of algorithm, they applied SVM classifier to distinguish non-ring structure (NRS) and organic ring structure (ORS) symbols, and then at the second stage, HMM method was employed for fine recognition. The average recognition rate for this algorithm is 98.08%. Unfortunately, all existing researches have been concentrated on recognizing isolated digits; however