| The use of image reranking to boost retrieval performance has been found to be successful for simple queries. It is, however, less effective for complex queries due to the widened semantic gap. In addition, the importance of verb visual concepts has always been ignored in image retrieval for complex queries. Therefore, we take the verb visual concepts into account and then propose a new approach towards image reranking for complex queries. Given a complex query, we first detect verb-phrase and noun-phrases as visual concepts. Then we estimate the relevance scores from three layers, i.e., the sematic-level, visual-level as well as cross-modality level. Based on the relevance scores, we can obtain a new ranking list. The main work and innovations of this research are:1. The paper targets web images reranking for complex queries from the probabilistic perspective. This work unravels the unreliable initial ranking list problem of the existing image reranking approaches for complex queries.2. we propose an image reranking method for complex queries based on the detection of visual concepts. Meanwhile, it proposes a heutistic approach to detect noun-phrase and verb-phrase from complex queries instead of just treating individual terms as possible concepts.3.It proposes a heterogeneous probabilistic network to automatically estimate the relevance score of each image. This network comprises three subnetworks, each representing a layer of relationship, including:the underlying relationship among image pairs, the cross-modality relationship between visual concept and the visual concept, and the high-level semantic relationship between visual concept and the complex query. The three layers mutually reinforce each other to facilitate the estimation of relevance scores for new ranking list generation. This method does not depend on the initial ranking result, and the whole process is unsupervised, so it can effectively solve the problems existing in the previous reranking approaches. |
PDF Full Text Request