RAIL CORRUGATION DETECTION BY GABOR FILTERING Clelia Mandriota, Ettore Stella, Massimiliano Nitti, Nicola Ancona, Arcangelo Distante Istituto Elaborazione Segnali ed Immagini – C.N.R. Via Amendola, 166/5 – 70126 Bari (Italy) ABSTRACT Inspection of the rail state in order to detect defects is one of the basic tasks in railway maintenance. Rail defects exhibit different properties and are divided in various categories relating to the type and position of flaws on the rail. In this paper, we propose a technique, based on texture analysis of rail surface, to detect and classify a particular class of defects: corrugation. 1. INTRODUCTION Rail defects are due to several kinds of interrelated factors such as the type of rail, the construction conditions, the speed and/or frequency of trains using the rail. With growing of the high-speed traffic on the rail tracks, the current research is oriented to develop inspection system able to detect automatically the rail defects without a human operator. In the past, the presence of human operator, during periodical visual monitoring in the detection defect phase, made inspection system slow and expensive. As alternative special wagons were equipped with any kind of sensors for monitoring of railway components. Aim of this paper is to present a technique able to detect the presence/absence of a defect and automatically classify it. This rail inspection technique is based on analysis of textured rail surface. Texture analysis is important in many applications of classification, inspection and segmentation of images based on local spatial pattern of intensity. In the paper we propose a technique to detect rail corrugation (see Figure 1) that is a particular class of defects that produce an undulatory deformation on the head of the rail. In high speed train, the corrugation induces harmful vibrations on wheel and on its components, reducing the lifetime. In the literature, many proposed texture segmentation schemes are based on a filter bank model, followed by some energy measure. Objective of the energy measure is to evaluate the energy in the filter output in a local region. Filtering approaches include Law mask, ring/wedge filters, Gabor filter bank, Wavelet transform, quadrature mirror filters, etc.[5]. We use, for our aims, a set of Gabor filters. The choice is justified because these approximate the characteristics of certain cells in the visual cortex of some mammals [4]. Moreover, Gabor functions are useful because achieve the lower bounds of the uncertainty inequalities π 4 1 Δ Δ u x and π 4 1 Δ Δ v y , and achieve optimally joint resolution in space and spatial frequency [1]. After this phase, a feature vector is extracted for each filtered Gabor image evaluating mean and variance as local energy functions. In fact, the analysis of textured surface involves the identification of texture attributes which can be used for segmentation. The paper is organized as follows: section 2 describes the Gabor filters; section 3 describes the texture features and finally in section 4 experimental results are shown on real data. 2. GABOR FILTERS In spatial domain 2D complex Gabor filter is given by ( ) 1 - = j : ( ) ( ) ( ) x jF y x g y x h = π 2 exp , , (1) ( ) + - = 2 2 2 2 2 1 exp 2 1 , y x y x y x y x g σ σ σ πσ where ( ) ( ) θ θ θ θ cos sin , sin cos , y x y x y x + - + = are the rotated coordinates, x σ and y σ are the standard deviations of Gaussian envelope along the x and y directions, F and θ are frequency and orientation respectively [7]. We consider only circularly symmetric Gaussians then σ σ σ = = y x . Thus (1) is a complex sinusoidal grating modulated by a 2D gaussian function 626 0-7803-6725-1/01/$10.00 ©2001 IEEE