Gabor Features for Offline Arabic Handwriting Recognition Jin Chen ∗ Lehigh University Bethlehem, PA 18015 jic207@cse.lehigh.edu Huaigu Cao Raytheon BBN Technologies Cambridge, MA 02138 hcao@bbn.com Rohit Prasad Raytheon BBN Technologies Cambridge, MA 02138 rprasad@bbn.com Anurag Bhardwaj † University of Buffalo Amherst, NY 14260 ab94@buffalo.edu Prem Natarajan Raytheon BBN Technologies Cambridge, MA 02138 pnataraj@bbn.com ABSTRACT Many feature extraction approaches for off-line handwriting recognition (OHR) rely on accurate binarization of gray- level images. However, high-quality binarization of most real-world documents is extremely difficult due to varying characteristics of noises artifacts common in such documents. Unlike most of these features, Gabor features do not re- quire binarization of the document images, and thus are likely to be more robust to noises in document images. To demonstrate the efficacy of our proposed Gabor features, we perform subword recognition for off-line Arabic hand- written images using Support Vector Machines (SVM). We also compare the recognition performance with other bina- rization based features which have been proven to be effec- tive in capturing shape characteristics of handwritten Ara- bic subwords, such as GSC (a set of gradient, structure, and concavity features) and skeleton based Graph features. Our preliminary experimental results show that Gabor fea- tures outperform Graph features and are slightly better than GSC features for Arabic subword recognition. In addition, by combining Gabor and GSC features, we obtain a signifi- cant reduction in classification error rate over using GSC or Gabor features alone. Categories and Subject Descriptors I.5 [Pattern Recognition]: Miscellaneous General Terms Algorithms ∗ Jin Chen performed this work during his summer internship at Raytheon BBN Technologies † Anurag Bhardwaj performed this work during his summer internship at Raytheon BBN Technologies Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. DAS ’10, June 9-11, 2010, Boston, MA, USA Copyright 2010 ACM 978-1-60558-773-8/10/06 ...$10.00 Keywords Arabic handwriting recognition, Feature extraction, Gabor filtering 1. INTRODUCTION There has been extensive work in the literature regarding features extraction approaches in the off-line Arabic hand- writing recognition. Many of these methods require high quality binarization of the document images [14, 4] which is difficult due to varying characteristics of noisy artifacts com- mon in such documents [16, 7]. In addition, large amount of gray-level information is lost during binarization. Therefore, features that are extracted from the original gray-level im- ages should be useful to discriminate handwritten character shapes. Gabor filters, which operate directly on gray-level images, have several advantages [12]. First, Gabor features have been used for capturing local information in both spatial and frequency domains from images, as opposed to other gobal techniques such as Fourier Transforms. Second, Gabor filters are orientation specific. This property allows us to analyze stroke directions in the handwriting. Third, the filtering output is robust to various noises since Gabor filters use information from all pixels in the kernel. Liu, et.al., compared Gabor features and gradient features for the task of off-line character recognition [11]. From their experiments, Gabor features were implemented in a wavelet framework where scaling factors were empirically designed. From experiments involving three databases: two handwritten digit datasets (MNIST and CENPARMI) and a printed Japanese character database (JCDB), the authors found that Gabor features outperformed gradient features at high recognition rates (> 95%) in the MNIST and the JCDB. Haboubi, et.al. [9] conducted a comparative study on four different methods of feature sets for Arabic handwritten script, i.e., structural approach, Fourier descriptors, pixel representation, and Gabor features. Surprisingly, their ex- perimental results showed that Gabor features did not per- form as well as the other three: with a database of 16,107 images, the classification rate for Gabor features was 19.3%,