UNSUPERVISED TEXTURE SEGMENTATION USING

-This paper presents a texture segmentation algorithm inspired by the multi-channel filtering theory for visual information processing in the early stages of human visual system. The channels are characterized by a bank of Gabor filters that nearly uniformly covers the spatial-frequency domain, and a systematic filter selection scheme is proposed, which is based on reconstruction of the input image from the filtered images. Texture features are obtained by subjecting each (selected) filtered image to a nonlinear transformation and computing a measure of "energy" in a window around each pixel. A square-error clustering algorithm is then used to integrate the feature images and produce a segmentation. A simple procedure to incorporate spatial information in the clustering process is proposed. A relative index is used to estimate the "'true" number of texture categories. Texture segmentation Multi-channel filtering Clustering Clustering index Gabor filters Wavelet transform I . I N T R O D U C T I O N Image segmentation is a difficult yet very important task in many image analysis or computer vision applications. Differences in the mean gray level or in color in small neighborhoods alone are not always sufficient for image segmentation. Rather, one has to rely on differences in the spatial arrangement of gray values of neighboring pixels-that is, on differences in texture. The problem of segmenting an image based on textural cues is referred to as the texture segmentation problem. Texture segmentation involves identifying regions with "uniform" textures in a given image. Appropriate measures of texture are needed in order to decide whether a given region has uniform texture. Sklansky (o has suggested the following definition of texture which is appropriate in the segmentation context: "A region in an image has a constant texture if a set of local statistics or other local properties of the picture are constant, slowly varying, or approximately periodic". Texture, therefore, has both local and global connotations--i t is characterized by invariance of certain local measures or properties over an image region. The diversity of natural and artificial textures makes it impossible to give a universal definition of texture. A large number of techniques for analyzing image texture has been proposed in the past two decades/2,3) In this paper, we focus on a particular approach to texture analysis which is referred to as ° This work was supported in part by the National Science Foundation infrastructure grant CDA-8806599, and by a grant from E. I. Du Pont De Nemours & Company Inc. the multi-channel filtering approach. This approach is inspired by a multi-channel filtering theory for processing visual information in the early stages of the human visual system. First proposed by Campbell and Robson (4) the theory holds that the visual system decomposes the retinal image into a number of filtered images, each of which contains intensity variations over a narrow range of frequency (size) and orientation. The psychophysical experiments that suggested such a decomposition used various grating patterns as stimuli and were based on adaptation techniques. I~l Subsequent psychophysiological experiments provided additional evidence supporting the theory. De Valois et al. ,(5) for example, recorded the response of simple cells in the visual cortex of the Macaque monkey to sinusoidal gratings with different frequencies and orientations. It was observed that each cell responds to a narrow range of frequency and orientation only. Therefore, it appears that there are mechanisms in the visual cortex of mammals that are tuned to combinations of frequency and orientation in a narrow range. These mechanisms are often referred to as channels, and are appropriately interpreted as band-pass filters. The multi-channel filtering approach to texture analysis is intuitively appealing because it allows us to exploit differences in dominant sizes and orientations of different textures. Today, the need for a multi-resolution approach to texture analysis is well recognized. While other approaches to texture analysis have had to be extended to accommodate this paradigm, the multi-channel filtering approach, is inherently multi-resolutional. Another important