Relevance Feedback Based on Feature Discreteness for Image Content Retrieval

Abstract

We propose a new relevance feedback approach to achieving high system accuracy for image content retrieval and high flexibility in selecting features of interest. Our system is divided into two major phases, namely index map construction and query comparison. In the phase of index map construction, ten MPEG-7 descriptors and six Tamura features are extracted from each database image. For each color or texture feature, database images are clustered by a self-organizing map and then an index map is constructed. In the phase of query comparison, a user can submit a query image and select features of interest. To search similar images from the database, we propose an image voting scheme. In each index map, database images that belong to the same cluster with the query are cast a vote. For each database image, its vote in each map is accumulated and the sum is regarded as the similarity to the query. If the retrieval result is not satisfied, the user can select relevant images as a new query. To improve the retrieval result, we propose a vote re-weighting scheme based on feature discreteness of relevant images. The database images that most similar to relevant images can be retrieved. Experimental results reveal the effectiveness of our approach.