Learning similarity for texture image retrieval

Abstract

A novel algorithm is proposed to learn pattern similarities for texture image retrieval. Similar patterns in different texture classes are grouped into a cluster in the feature space. Each cluster is isolated from others by an enclosed boundary, which is represented by several support vectors and their weights obtained from a statistical learning algorithm called support vector machine (SVM). The signed distance of a pattern to the boundary is used to measure its similarity. Furthermore, the patterns of different classes within each cluster are separated by several sub-boundaries, which are also learned by the SVMs. The signed distances of the similar patterns to a particular sub-boundary associated with the query image are used for ranking these patterns. Experimental results on the Brodatz texture database indicate that the new method performs significantly better than the traditional Euclidean distance based approach.