A Relevance Feedback Method to Trademark Retrieval Based on SVM

Relevance feedback is a good method for the semantic gap between the low-level similarity and the high-level user's query in content-based image retrieval. It interactively asks user whether certain proposed images and the query output are relevant or not. In this paper we propose the use of a support vector machines for conducting effective relevance feedback for trademark retrieval. The algorithm selects the Tamura textures feature which consistent with human vision perception and the low-level feature of images. Experimental results show that it achieves significantly higher search accuracy after just three or four rounds of relevance feedback.