Arabic (Indian) Numeral Handwritten Recognition Using Angular Radial Transform

In this paper, an Arabic (Indian) numeral handwritten recognition method is presented based on angular radial transform. The angular radial transform is considered as a global features extraction descriptor in order to provide distinct and rotation invariant features about the images of Arabic numeral handwritten. Also, in this, paper the performance of both angular transform and radial transform is investigated and compared. Hellinger distance measure is adopted in the classification stage to compute the distance between the test and training Arabic numeral handwritten images. The extensive experiments indicate that the proposed approach achieved a high recognition rate of 96.74% which is better of recognition rates achieved using its counterpart's angular and radial transforms which achieved 91.34% and 87.10% respectively. Also, they indicated that the performance of angular transforms is outperforms the performance of radial transform. Furthermore, observed that the proposed method is rotation invariant.


Introduction
The writing is a natural process where the brain sends a nerve impulse to the hand, to start writing which means that the handwritten script is differs from person to another as other human features such fingerprints, pinna, and facial expressions [1].Therefore, handwritten recognition can be used to distinguish between the people.Handwritten recognition can be classified into two types which are on-line and off-line handwritten recognition [2,3].On-line handwritten recognition means that the system has the ability to recognize the handwritten in real time.On the other hand, in off-line handwritten recognition, the recognition process is achieved indirectly.Numeral handwritten recognition has received more attention in the last years due to its wide applications in different fields such as criminal evidence, office computerization, extract distinct features from Arabic numeral handwritten images.Furthermore, these features are rotation invariant.In order to compute the distance between the extracted features of the test and training of handwritten numeral images, the Hellinger distains measure has been used.
Extensive experiments have been carried out to evaluate the accuracy of the proposed work.
The rest of the paper has been adjusted as follows: Section 2 presented an overview of related work.The ART is discussed in Section 3. Section 4 explained the used similarity measures.
The proposed method is given in Section 5.The experimental analysis is presented in Section 6. Section 7 clarifies the conclusions.

Related work
The main challenges in handwritten recognition process are the distortions and enormous variability of these patterns [8] Therefore, any successful handwritten recognition system required an active and accurate feature extraction technique that can provide distinct features which can be used effectively to distinguish between different numeral handwritten images.introduced an Arabic (Indian) numeral recognition method, based on centralized moments.In this method the features of numeral images are extracted by estimating the centralized moments for vertical and horizontal.The work in [10] presented by Ouafae et.al.proposed a new handwritten numeral recognition system using Characteristics Loci(CL).In this method each numeral image is divided into four portions, and then the CL derived from each portion of the image.This work adopted two type of the classifiers in the classification stage, which are multilayer perception and k-nearest neighbor's classifiers.Meng Shia et.al.[11] presented a handwritten numeral recognition technique using curvature and gradient transformation.In this work there are three main processes are achieved in order to extract the features of handwritten numeral images, these processes are curvature and gradient calculation, feature vectors generation, and dimensionality reduction.Furthermore, in this paper, the experiments carried out on three standard handwritten numeral databases, namely, IPTP CDROM1, NIST SD3 and SD7.Kathirvalava kumar Thangairulappan and Palaniappan Rathinasamy [12] presented handwritten Arabic numeral classification base on neural network.In this method the image of handwritten numerals are converted into the matrix, then the size of each matrix reduced in half by using logical OR operation.Vikas J. Dongre and Vijay H. Mankar [13]

Angular radial transform
The ART can be defined in polar coordinates on a unit disc based on complex orthogonal sinusoidal basis functions.The ART can be defined in mathematical form as follows [18]: where Eq.( 2) can be divided into separate parts.One is the radial part which is only dependent on the distance from the center and defined in terms of  only.This is called the radial part of the basis function.Another part is called as angular part and this is dependent on the angle from the x-axis and defined in terms of  .The radial part of the basis function is where p is a non negative integer, q is an integer, It can be defined as follows [19]     4 presented the average of recognition rate achieved using the proposed method and its counterparts (AT and RT).The existing results refer that the proposed method achieved 96.74% recognition rate which is better than 91.34% and 87.10% which achieved by using AT and RT respectively.Also, it observed that the performance of AT is better than the performance of RT. , while, the training numeral images are taken without rotation angle.Figure 3 shows original Arabic (Indian) numeral image and its rotation version using different rotation angles.Table 5 presented the results of the experiments which carried out in this section.It observed that the recognition rates achieved using rotate test Arabic (Indian) numeral images by angles Also, it required an accurate classifier to compute the exact distance between the feature vectors of the test and database numeral handwritten images.There are numerous numeral handwritten recognition have been existed based on different feature extraction and classifier techniques.Sabri A. Mahmodi and Marwan.Abu-Amara [9] proposed Arabic (Indian) numeral handwritten recognition method based on radon and Fourier transforms.In this work, the Fourier transform is applied on each projection of radon transform for each handwritten to construct feature vectors of the training and test Arabic numeral images.In the classification stage, the researchers consider three types of classifiers which are hidden Markov model classifier, Vol: 13 No:2 , April 2017 DOI: http://dx.doi.org/10.24237/djps.1302.207BP-ISSN: 2222-8373 E-ISSN: 2518-9255 nearest mean classifier, and K-Nearest neighbor classifier.Mohamed H. Ghaleb et.al.[8] proposed Devnagari numeral recognition method based on geometrical features.Statistical combination classifier utilized to distinguish between the training and test Devnagari Handwritten numeral images.Gita Sinha and Dr. Jitendra kumar [14] presented offline Arabic numeral handwritten recognition techniques using Zone based feature extraction method.In this method, the researchers consider three feature extraction techniques namely, zone centroid zone (ZCZ), image centroid zone (ICZ), and combined feature extraction method consist of ICZ+ ZCZ.In this work the support vector machine classifier is utilizes in the classification stage.F.A. Al-Omari and O. Al-Jarrah [15] presented online Indian numeral handwritten recognition system based on geometric features.In this system, the authors clarified that the proposed system is a rotation, scaling, and translation invariant.The probabilistic neural network is adopted to Vol: 13 No:2 , April 2017 DOI: http://dx.doi.org/10.24237/djps.1302.207BP-ISSN: 2222-8373 E-ISSN: 2518-9255 distinguish between the training and test Indian numeral handwritten images.The method in [16] proposed by Javad Sadri et.al.presented isolated Arabic -Persian handwritten digits.In this method, the features have been extracted by considering each digit from four views which are left, right, bottom, and top.The distance between extracted features is computed using support vector machines.Md Sohail Siddique and Ayatullah Faruk Molla [17] proposed Arabic and Urdu numerals handwritten recognition based on the digit variations.
No:2 , April 2017 DOI: http://dx.doi.org/10.24237/djps.1302.207BP-ISSN: 2222-8373 E-ISSN: 2518-9255 Hellinger distance measure has been considered to compute the distance between the training and test numeral handwritten images.Hellinger distance measure estimate the deviation between two objects.The resulted values of using Hellinger distance measure are within 0 and 1.

2. Evaluate the rotation invariant property of the proposed method In
this section, the rotation invariant trait of the proposed method has been demonstrated through a set of experiments using different rotation angles.For this purpose, we have rotated the numeral test images of the dataset used in Section 6.1 by