The USC Andrew and Erna Viterbi School of Engineering USC Signal and Image Processing Institute USC Ming Hsieh Department of Electrical Engineering University of Southern California

Technical Report USC-SIPI-111

“Hierarchical Texture Segmentation Using Multiple Resolution Operators”

by Allan G. Weber

July 1987

Texture information can be used as a basis for segmenting an image into regions. Image texture is basically a local area phenomenon that is sensitive to the size of the area. What appears as a non-textured area at one resolution level can appear as a region with distinctive texture at a different resolution. The performance of texture segmentation schemes is often highly dependent on the size of the local area operator used to generate the classification features. The size of the operators has a major impact on the performance near the boundaries between texture regions. Features based on large operators perform better overall but are highly affected by the mixing of class statistics when the operator overlaps more than texture, such as near the texture boundaries. Features based on small operators show poorer overall performance but are more likely to maintain an acceptable performance level in the boundary areas. The trade-off is between statistical accuracy of the classification versus the final accuracy of the texture region boundaries. The problem being studied here is how to combine information from texture classifiers operating at different resolutions into a segmentation process that gives acceptable performance in all areas of an image.

In this study, the nature of the mixing problem is examined and a solution is proposed based on using multiple resolution features in a hierarchical decision process. A key component of the solution is an analysis of the image data prior to performing any classification. From this analysis, we determine the expected location in the feature space of the mixture points. Three different methods of isolating the mixture points in the feature space are proposed and tested. During the initial classification phase, image points that are within the mixture areas are left unclassified. Spatial information is incorporated into the segmentation process by the use of a local area cohesion operation. The final segmentation is based on a hierarchical decision process that uses the classification choice at both resolutions and the spatial cohesion data. Several decision processes are tested that use the information in different ways.

To download the report in PDF format click here: USC-SIPI-111.pdf (19.9Mb)