Abstract

A requirement of any vision system is to segment the image into regions having a set of common characteristics. In many applications, this common characteristic is texture. Texture is a higher order image property, and depends on the statistics of pixels in a local neighborhood. To perform segmentation, the regions of homogeneous higher order statistics must be identified and the spatial boundaries where statistical properties change must be located. These local statistics are derived from measurements within a window whose dimensions are subject to a conflicting set of requirements. Using a spatially large window produces better statistical accuracy in homogenous regions having stationary statistics, but it averages over boundaries between regions of different texture, leading to many incorrectly segmented pixels. Using a spatially small averaging window produces more accurate classifications near texture boundaries, but has poor statistical accuracy compared to the large windows. The optimum window sizes to be used are not known a priori, thus we use a hierarchy of averaging windows having a range of sizes and resolutions. The overall goal is to improve the overall segmentation quality over that obtained at a single resolution.

© 1985 Optical Society of America