Abstract
Region-based approaches and relevance feedback have been indispensable issues in content-based image retrieval (CBIR). Region-based representation combines both local information and their spatial organization so as to provide better image representation and matching criterion. Relevance feedback allows users to rate retrieved images and refine the retrieved results interactively. Current relevance feedback methods are mainly divided into intrasession and intersession learning, depending on whether or not the learned information from historical query sessions is accumulated to subsequent query sessions. This study addresses region-based issues for both intrasession and intersession learning. The proposed intrasession learning technique is a generalized Bayesian framework which incorporates a time-varying user model. The user model includes a target query to specify the user’s ideal query and a user conception to adjust the time-varying matching criterion. We include spatially adjacent relationship to estimate the region correspondence between images for better image matching criterion. In addition, we also propose to update the target query as well as the user conception in region level based on the estimated region correspondence. For intersession learning, we aim to infer the hidden semantic space in region level by accumulating the knowledge learned from previous query sessions. The main idea is that a region in the query possesses a hidden semantic concept, and hence a query session will generate several concepts in our work. We initialize the hidden semantic space based on a series of query sessions. With the constructed hidden semantic space, we then perform retrieval and keep accumulating newly learned concepts into the hidden semantic space. Since the hidden semantic space may contain inconsistency, overlapping or mislabeled concepts, we employ a dimension reduction technique in region level to construct a more compact space. Experiments demonstrate that the proposed intrasession learning method combined with time-varying user model and region-based representation achieves satisfactory results. The results also show that our intersession learning method based on the inferred hidden semantic space further improves the retrieval accuracy, and the proposed dimension reduction technique removes redundant hidden semantic concepts effectively.