Image Retrieval With High-Level Semantics

Efficient Information retrieval has been a crucial issue in the era of worldwide web. The history of Content-based Image Retrieval (CBIR) is longer than30years. However, there still remains some problems unsolved. Two of the most well-known problems are the scalability in large-scale data and the assessability to visual semantics. In this paper, we give an comprehensive survey on these problems and propose innovative solutions.We propose an enhanced codebook training algorithm (CSLC) and an efficient geometric verification scheme (BVP). CSLC uses principal component analysis to carefully select initial points, and leader clustering to reduce noisy samples. BVP encodes local geometry of features and enables simultaneously geometric consistency checking in searching.We also initiatively introduce click-data to bridge the semantic gap in CBIR. Click-data is a perfect source for learning visual semantics and we propose a semantic re-ranking method based on k-nearest neighbor scoring and metric learning. The proposed re-ranking method can adapt to both example-based and keyword-based image retrieval.Exhaustive experiments are carried out on multiple open dataset, including oxford, paris, flickr100k and clickture. The results show that significant improvement is achieved by the proposed methods, in comparison to existing methods. We also introduce our outstanding performance in TRECTVID INS2013...