Learning user preference in a personalized CBIR system

A new approach for learning user preference in a personalized content-based image retrieval (CBIR) system is proposed in this study. This approach provides users with textual descriptions, visual examples, and relevance feedbacks to find target images. The user query can be expressed by syntactic rules and semantic rules. To build a personalized CBIR system, two problems should be overcome in advance, including the semantic gap and the human perception subjectivity. In this study, the semantic gap is bridged through linguistic term sets, which are represented as fuzzy membership functions. The human perception subjectivity is modelled from relevance feedbacks through profile updating and feature re-weighting algorithms. The user preference is stored in a personal profile for further retrieval. Experimental results support the effectiveness of the proposed approach.