Novel unsupervised multiresolution texture segmentation approach

Abstract

Image texture plays a vital role in the segmentation process. A novel unsupervised segmentation approach based on multiresolution cooperative texture model computation is developed. The multiresolution segmentation approach is based on the observation that the human visual system utilizes relatively `global' information about an image in conjunction with `local' information to reach segmentation decisions. The texture model developed is based on sets of gray level co-occurrence matrices rather than measures extracted from them. The concept of multiresolution associated region (MAR) is developed for pyramid schemes. The other algorithmic constituents for the segmentation scheme such as normalized match distances between texture models, region homogeneity criteria with extensions to MARs, are systematically developed. The MAR aggregation rule is utilized to perform segmentation decisions at the base resolution level. The segmentation strategy was tested extensively on natural texture mosaics as well as on real scenes and the results are analytically presented. An important observation was that smaller texture models at multiple resolutions performed better than a very large texture model at single resolution.