Host: Japan Society for Fuzzy Theory and Intelligent Informatics (SOFT)
Visual key image retrieval system has been proposed, which depends on a part of images. In the conventional research, the similarity used for constructing clusters was defined by linear combination of Euclid distance of each feature and the retrieval precision is 10%. Linear sum has been used without being discussed enough about the distance function used to calculate the degree of similarity and the weight has been decided empirically. Similarity suitable for person's sensibility can be constructed by changing the combination of the feature vector and the operator more appropriately, and the distance function which is better than conventional techniques can be obtained. However, it takes a great amount of time to search the combination of operators suitable for person's sense, and global optimum is difficult to be derived from among the combination of huge number of operators. In this research, genetic programming is used to search the suitable combination of the aggregation operators and feature vector. Compared with conventional technique, the recall ratio is 8% decrease and the relevance ratio is 19% increase and the F-value increase by 7%.
