Relevance feedback in content-based image retrieval system by selective region growing in the feature space

This paper proposes a relevance feedback algorithm for the content-based image retrieval system. In the conventional algorithms, the weights of feature vectors are adjusted based on the user's feedback, which warps the match region from the hyper-sphere to hyper-ellipsoidal shape. That is, the axis grows into the direction that covers more relevant images in the feature space. The proposed algorithm is not based on the adjustment of the weights, but on the generation of new match region based on the user's feedback. Specifically, new spheres centered at the relevant images are generated, the radius of which are determined by the number of neighboring relevant and irrelevant images. The overall match region is the union of all the spheres generated and modified at each iteration of feedback process. As a result, the match region grows in a bubble shape into the direction where there are more relevant images. The resulting match region can cover arbitrarily shaped clusters whereas the weight updating approach can cover only hyper-ellipsoidal region. We also propose a data structure that keeps the history of past searches, for more rapid expansion of match region.