Spatial Scene Matching Based on Multilevel Relevance Feedback

Existing spatial scene matching methods try to use low level features such as local geome-trical shape descriptors, binary spatial relations of cardinal directions or topology and so on to meet with the high level description demands from users, and lack the feedback mechanism of user subjective concepts. This paper applies the feedback relevance mechanism to spatial scene matching, analyzes the effects after studying users' needs, and then evaluates the relevance of matching results to the desired spatial scene. According to the feedback, the retrieval model updates the weights of descriptors to simulate user's perception, as a result, the retrieval fits the users' requirements much better. The experiment verifies the efficiency and feasibility by users survey and weights convergence. Experimental result shows that spatial scene matching based on multilevel relevance feedback has a high user subjectivity and accords with human's cognizance. In addition, users could obtain satisfactory result by two or three iterations.