A hierarchical manifold subgraph ranking system for content-based image retrieval

We present a novel hierarchical manifold subgraph ranking system for content-based image retrieval (CBIR). The proposed CBIR system is capable of searching a large scale imagery database via its hierarchical structure. To achieve this scalability, we first apply the SVM-based learning mechanism to construct users´ relevance feedback groups and extract high-level semantic features for each image. We then build two-layer manifold subgraphs by incorporating both visual and semantic similarity into the second-layer manifold subgraphs to achieve more meaningful structures for the image space. Finally, a relevance vector is created for each subgraph in the second-layer manifold subgraph by assigning initial scores from the first-layer manifold subgraph. These asymmetric vectors are further used to propagate relevance scores of labeled images to unlabeled images via hierarchical manifold subgraphs. Our extensive experimental results demonstrate the proposed system achieves the best retrieval accuracy when comparing with two manifold-based and five state-of-the-art CBIR systems in the context of correct and erroneous users´ feedback.