Scene image retrieval via re-ranking semantic and packed dense interestpoints

In this paper, we propose a novel method for scene image retrieval in which the semantic meaning of an image and a new low-level feature are combined. The fluid nature of scene images makes learning semantics essential in our retrieval task. Compared to a general image, a scene image contains large regions of low contrast, which makes it difficult for a method to extract features that has good coverage of the entire image and assurance of relatively high repeatability. Given a scene image as a query, a collection of images is first retrieved by some search engines based on the images' semantic meanings. The candidate images are re-ranked by adapting an asymmetric piece-to-image matching scheme based on their visual similarities with the query image, using its visual signature consists of some packed dense interest points. Our method is evaluated on an Outdoor Scene Recognition (OSR) dataset and an NUS-WIDE dataset. It has demonstrated the improvements of our method over other conventional approaches.