An Adaptive Fuzzy Clustering Method for Query-by-Multiple-Example Image Retrieval

Content-based image retrieval (CBIR) methods usually adopt color, texture and structure as image feature vector. Recent research indicates that since these features are not exactly associated with image semantic meaning, query-by-one-example (QBOE), which means to query with only one image, usually is insufficient to achieve good performance. Thus, query-by-multiple-examples (QBME) methods are introduced and applied in many content-based image retrieval systems. However, how to maximize major features and minimize minor ones of these inputs while matching could influence retrieval results significantly. Some previous researchers divided input images to groups according to their relevance to desired image class. They kept important features of relevant group and discard obvious features of irrelevant group while matching. These methods did improve the retrieval results but have defects in some situations. This paper presents an improved QBME system which optimizes retrieval result. It adopts clustering method to cope with the defects of previous methods. Uniquely, to decrease the complexity of user input and reduce user-computer interaction, a modified fuzzy clustering algorithm will be introduced. It not only presents good performance but also requires no parameters from users. In this article, both the defects of previous methods and new methods to cope with them will be explained in detail.