Improving Image Retrieval with Semantic Classification Using Relevance Feedback

In this paper, we investigate the combination of image semantic classification with content-based image retrieval. A flexible scheme is proposed to take advantage of image classification, which may be obtained manually or automatically, to enhance image retrieval. In this scheme, a semantic feature vector is composed for an image based on its class membership information, and is combined with low-level features in image retrieval. Relevance feedback techniques are also used to adjust both the semantic feature and low-level features of the query in order to better reflect the user’s intention. Experimental results on a collection of 10,000 images with manual classification demonstrate the effectiveness of the proposed method.