A Systematic Approach to Semantics-Based Image Retrieval and Organization Using Thesaurus

Semantics-based retrieval capability is greatly desirable for large-scale image collections. This paper proposes a novel approach to semantics-based image retrieval and organization using thesaurus, which addresses the representation, acquisition, and utilization of image semantics in a systematic and integrated manner. Firstly, a semantic network is constructed from thesaurus, which represents the semantic indexes of images using a group of interrelated semantic concepts. A learning strategy is suggested to acquire the semantic indexes from user interactions (including feedbacks) in a progressive and semi-automatic way. Based on the semantic network, a semantic similarity measure is formulated and integrated with visual features to compose an integrated image retrieval algorithm. A set of experiments conducted on real-world images has demonstrated the efficiency and effectiveness of this algorithm compared with existing retrieval algorithms. Furthermore, we propose a method to generate a dynamic image classification that facilitates flexible and convenient user interactions. Keyword: image retrieval, semantic index, thesaurus, image classification, learning strategy 1This paper extends the idea of a short paper presented in the proceedings of ICME 2001 [28], August 2001, Japan.