Texture characterization using a curvelet based descriptor

Abstract

Feature extraction from images is a key issue in image classification, image representation and content based image retrieval. This paper introduces a new image descriptor, based on the curvelet transform. The proposed descriptor captures edge information from the statistical pattern of the curvelet coefficients in natural images. The image is mapped to the curvelet space and each subband is used for establishing the parameters of a statistical model which captures the subband marginal distributions as well as the dependencies across scales and orientations of the curvelet. Finally, the Kullback-Leibler distance between the statistical parameters is used to measure the distance between images. We demonstrate the effectiveness of the proposed descriptor by classifying a set of texture images, and with a simple nearest neighbour classifier we obtained an accuracy rate of 87%. © 2009 Springer-Verlag Berlin Heidelberg.