Abstract:

In order to overcome the inherent asymmetry and the small sample size of relevance feedback （ RF）, a RF algorithm of image retrieval is proposed based on the modified multi-class fuzzy support vector machines （ FSVMs）. In this algorithm, the RF is considered as a multi-class classification problem between one relevance class and several irrelevance classes, and the original membership function of FSVMs is modified to avoid negative values. Moreover, the conventional constrained random selection method is extended to a multi-class case, and a memory marking method is used to lighten the burden of multi-class marking and to decrease the classification error. Experimental results demonstrate that the proposed algorithm helps to obtain satisfying retrieval results with less feedback times.

Key words: multi-class fuzzy support vector machine, multi-class constrained random selection, memory marking, image retrieval, relevance feedback