* Corresponding author. Tel.: #91-33-577-4906; fax: #91- 33-577-6680. E-mail addresses: res9522@isical.ac.in (M. Acharyya), malay@isical.ac.in (M. K. Kundu). Signal Processing 81 (2001) 1337}1356 An adaptive approach to unsupervised texture segmentation using M-Band wavelet transform Mausumi Acharyya, Malay K. Kundu* Machine Intelligence Unit, Indian Statistical Institute, 203 Barrackpore Trunk Road, Calcutta 700 035, India Received 9 December 1999; received in revised form 5 December 2000 Abstract The M-band wavelet decomposition, which is a direct generalization of the standard 2-band wavelet decomposition is applied to the problem of an unsupervised segmentation of two texture images. Orthogonal and linear phase M-band wavelet transform is used to decompose the image into MM channels. Various combinations of these bandpass sections are taken to obtain di!erent scales and orientations in the frequency plane. Texture features are obtained by subjecting each bandpass section to a nonlinear transformation and computing the measure of energy in a window around each pixel of the "ltered texture images. The window size in turn is adaptively selected depending on the frequency content of the images. Unsupervised texture segmentation is obtained by simple K-means clustering. Statistical tests are used to evaluate the average performance of features extracted from the decomposed subbands.  2001 Elsevier Science B.V. All rights reserved. Keywords: M-band wavelets; Texture segmentation; Feature extraction; Multiscale representation 1. Introduction Most natural surfaces exhibit texture. The char- acterization of texture plays an important part of many computer vision system. Texture analysis has wide range of applications like medical diagnosis, content-based-image retrieval, satellite imaging and many others. Image segmentation is a di$cult yet very important task in image analysis and many computer vision applications. The problem of seg- menting an image into meaningful regions based on textural cue is referred to as texture segmentation problem. A large number of techniques for ana- lyzing textured image have been proposed in the past [13] and in a recent review, Tuceryan and Jain [34] have discussed some of those techniques. In this paper we focus on a particular approach to texture (image) analysis which is referred to as multichannel "ltering approach. This approach for texture analysis is intuitively appealing because it allows us to exploit di!erences in dominant sizes and orientations of di!erent textures. In several papers the successful applications of multichannel "ltering for texture segmentation were reported [10,16,3] using various "ltering techniques, such as isotropic "lters [5] discrete cosine transform (DCT) [23] and Gabor "lters. The reason for the popular- ity of Gabor "lters is due to their joint optimum 0165-1684/01/$-see front matter  2001 Elsevier Science B.V. All rights reserved. PII:S0165-1684(00)00278-4