| In this thesis we are proposing a new content-based image retrieval (CBIR) prototype that combines the advantages of relevance feedback and image database categorization. We have developed a novel algorithm, called Semi-Supervised Clustering and Attribute Discrimination (S-SCAD), that performs clustering and feature weighting simultaneously and can incorporate partial supervision information. The feature weighting is needed to address the fact that images of the same concept tend to share one feature similarity more than another. While supervision provides a correction for automated clustering based on visual features. This supervision information, extracted from the user's feedback through visual exploration and interaction, is used to refine the clusters' distributions and their feature relevance weights. The cluster dependent feature weights offer two main advantages. First, they guide the clustering process to partition the data into more meaningful categories. Second, they are used in the retrieval phase to adapt the similarity measures to different categories. The feedback information is encoded as a set of constraints on which images should or should not reside in the same cluster. Thus, it does not depend on the query image explicitly. Consequently, partial supervision information from different query sessions could be saved, accumulated, and used to continuously refine the image categories and their feature weights in multiple neighborhoods of the feature space. The proposed CBIR and clustering algorithm are evaluated on two collections of images. The results are visualized and analyzed quantitatively. |
