Color Texture Recognition in Video Sequences using Wavelet Covariance Features and Support Vector Machines D. K. Iakovidis 1 , D. E. Maroulis 1 , S. A. Karkanis 2 , I. N. Flaounas 1 1 RealTime Systems & Image Analysis Group, Department of Informatics and Telecommunications, University of Athens, Panepistimiopolis, Illisia, 15784 Athens, Greece (rtsimage@di.uoa.gr) 2 Technological Educational Institute of Lamia, Dept. of Informatics and Computer Technology, 3rd klm Old National Road, 35100 Lamia, Greece (sk@teilam.gr) Abstract This paper pertains to the recognition of textural regions for color video analysis. The proposed scheme uses the covariance of 2 nd -order statistics on the wavelet domain, between the different color channels of the video frames. These features, named as Color Wavelet Covariance (CWC), are used as color textural descriptors. A Support Vector Machine was chosen for the classification of the CWC feature vectors. Experiments were conducted using both animated Vistex texture mosaics and standard video clips. The estimated average accuracy ranged from 90% to 97%. The results show that the proposed methodology could efficiently be used in various multimedia applications as a complete supervised color texture recognition system. 1. Introduction Video sequence analysis is an arising research area, which becomes essential as multimedia applications enter in our everyday life. The increase of the computational power of modern workstations has made feasible the application of complicated image analysis techniques on video frames. Such techniques usually exploit color and texture, both fundamental properties of the visible surfaces. Significant research effort has concentrated to the mathematical representation of color and texture for video sequence analysis. State of the art applications exploiting these properties include object tracking [1], face detection and recognition systems [2], tumor detection in endoscopic video [3] and content indexing [4]. Recent studies in color texture analysis have considered the use of perceptual approaches [5], the use of chromaticity moments [6], the derivation of textural information from luminance channel along with pure chrominance features as well as the processing of each color channel separately, by applying gray-level texture analysis techniques [7]. Other approaches exploit the interdependence of the existent textural information within the different channels of a color image, usually captured by means of correlation. On this direction Van de Wouwer et al [8] achieved high classification rates using correlation signatures estimated from the wavelet coefficients of color images. Paschos [9] proposed a set of discriminative and robust chromatic correlation features using directional histograms. Vandebroucke et al [10] exploited the correlation of 1st order statistical features between the different color channels for unsupervised soccer image segmentation and Al-Rawi et al [11] proposed Zernike moments of correlation and covariance functions for illumination invariant color texture recognition. Frame Acquisition Feature Extraction Classification Post Proccessing Video Signal Input Output Frame Acquisition Feature Extraction Classification Post Proccessing Video Signal Input Output Figure 1. Color texture recognition scheme