Semantic clusters based manifold ranking for image retrieval
Ran Chang, Xiaojun Qi · 2011
We propose a novel weighted manifold-ranking based image retrieval method to improve the effectiveness of traditional manifold methods. Specifically, we apply the SVM-based relevance feedback technique to create semantic clusters for computing the reliability score of each database image. We then incorporate the reliability scores into the affinity matrix to construct a weighted manifold structure. We finally create an asymmetric relevance vector to store users' positively and negatively labeled information. Our system ensures to propagate the labels in the relevance vector to the images with high reliability scores and discriminately spread the ranking scores of positive and negative images via the weighted manifold structure. Extensive experiments demonstrate our system outperforms the other manifold systems and SVM-based systems in the context of both correct and erroneous feedback.