Summary: | 碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 93 === Relevance feedback (RF) and region-based image retrieval (RBIR) are two widely used methods to enhance the performance of content-based image retrieval (CBIR) systems. In this paper, these two methods are combined. Rather than using only one positive feedback group, the proposed approach embeds the RF in RBIR with multiple positive and negative groups. In order to objectively assist the user in grouping the positive feedbacks, a heuristic pre-clustering result is automatically provided. Using these guiding clusters, the user can easily and subjectively re-group the feedbacks to express his/her particular interest. Furthermore, a region weighting scheme fitting the human visual perception is proposed to enhance the weighting importance of the region whose pixels are closer to the attention center. Finally, Group Biased Discriminant Analysis (GBDA) is modified and applied to the similarity measure between images based on these region-based relevance feedbacks.
|