適用於圖片搜尋系統之擴充式語意機制之研究

Abstract

One major limitation of content-based image retrieval is the semantic gap between low-level features of images and high-level concepts of human. This is because users usually prefer querying images by high-level concepts or semantics instead of low-level images features. In this paper, we present an enhancement semantic-based mechanism (ESBM) to solve the problem. The mechanism is mainly based on combining the concept of relevance feedback and information filtering. Our system considers color semantics such as pretty, cheerful, etc. as one high-level conceptual query. In addition, images which have been classified into some conceptual categories are available for keyword-based queries. The system first of all learns users’ queries and cluster users into four groups by their ages, occupations, interests and gender. After training, new query results will be based on the feedbacks and user’s groups. Experimental results show that ESBM is able to enhance retrieval effectiveness compared with traditional keyword-based image retrieval systems.