GABOR WAVELET BASED AUTOMATIC COIN CLASSSIFICATION Taraggy M. Ghanem 1,2 , Mohamed N. Moustafa 1 , Hussein I. Shahein 1 1-Computer and Systems Department, Faculty of Engineering, Ain Shams University, Cairo, Egypt 2- Faculty of Computer Science, Misr International University, Cairo, Egypt ABSTRACT We present an automatic coin classifier mainly depending on visual features. Our multistage system starts out by segmentation using circular Hough Transform, features extraction by two complementary cues and finally classification by simple nearest neighbor measure. Our features extraction process relies on rotation invariant edge orientation followed by Gabor wavelet convolution. Testing on the publicly available portion of a benchmark European coins database, we can correctly classify 93.5% and 98% of the coins using single face and double faces images respectively. We also show that our correct classification rate can reach 99.8% when adding the coin thickness measurement (which is available for this database). 1. INTRODUCTION Building an accurate automatic coin classifier is a task with great beneficial role in charity organizations, cultural heritage domain and financial institutions, for sorting heterogeneous coin collections (modern and historical) automatically, and for building automatic cash machines and currency counters. Over the last years many systems [1, 2, 3, 4, 8] were built serving this field, depending on visual features in addition to other sensor measurements like radius, thickness and weight and since 2006, a competition is organized annually, with a prize sponsored by the Muscle Network of Excellence to find the best automated coin classification algorithm to deal with large volumes of mixed coin collections collected by charitable organizing after changing the twelve European currencies to the Euro. Our system is a new approach that serves this goal and depends basically on visual features. The system in [1] bases on three rotation invariant features derived from edge information, while in [2] it bases on vector quantization and histogram modeling. In [3] the system bases on computing translational, rotational and illumination invariant features in the Eigen space while in [4] bases on collinear gradient vectors. In [8] a classification system of partially occluded coins bases on polar gradient orientations. Our database is a set of coins from the benchmark European coins database [7] , each class is referenced by an average image, and the testing coins are samples for each class differing in orientation and illumination effects. The proposed procedure is organized as follows: in section 2 we discuss the segmentation phase, section 3 presents the phase of features extraction, we applied two different methods, the first depends on the phase of gradient vectors [4] , and the second is applying the Gabor wavelets. The output of the first method is a set of the best nearest neighbors which is then used as an input to the second method. Section 4 describes the classification scheme, section 5 shows our experiments and results and finally section 6 summarizes our conclusions and future works. In our proposed system, the algorithms applied in the segmentation phase and in the first feature extraction method are referenced to [4], our main contribution is applying the Gabor wavelets on the best nearest neighbors computed by the first method and also depending on visual features only not on other sensor measurements of thickness as happened in [1, 3, 4 and 8] and working on inhomogeneous background and relatively larger number of classes than [1, 2]. Our proposed work is summarized in fig1. Fig 1. overview over the paper (Bold Italic represents the output of each phase) 2. SEGMENTATION Image segmentation is an essential preliminary step in any pattern recognition system. The main objective of the segmentation phase is to satisfy translational invariance. Our objects are the circular coins. In this phase we applied the Segmentation phase Polar Representation Compute 1 st feature Collinearity Measure Compute 2 nd feature Test Radius , First feature image Set of nearest neighbours Class membership Apply ranking procedure Subset of nearest Database of references coins Collinearity Measure Segmentation phase 305 978-1-4244-5654-3/09/$26.00 ©2009 IEEE ICIP 2009