A solution to the problem of segmentation near edges using adaptable class-specific representation

Accurate segmentation of pixels near edges is important in applications where exact shape and size is critical. Image sampling traditionally involves moving a sampling window of fixed shape across an image. Mismatches in the spatial frequency domain between templates and new images occur when the sampling window contains an edge and more than one true segment. This paper presents a novel algorithm, which adapts the shape of the sampling window locally, approximating to optimal class-specific representations. Unique representations of the same pixel for different segment classes are generated before evaluation by a set of classifiers. The algorithm is not specific to a particular type of classifier or encoding scheme. In this paper the algorithm is demonstrated by shelving that it produces accurate segmentation with minimal or no edge artefacts of artificial and natural colour images using LVQ classifiers