Image Classification And Retrieval Based On Bag-of-Words Model

With the rapid development of information in modern society, there is a large scale of images being produced in humans’ daily life. Due to the huge number of images, as well as the inherently complicated and abundant semantic information, it is being an extremely challenging and concerned problem in Computer Vision that how to completely describe image semantics, and how to category and retrieve images efficiently and accurately. The Bag-of-Words(BoW) model provides a promising way for the high-level semantic representation and has become one of the most popular framework for image classification and retrieval. In this thesis, we focused on the study of image classification and retrieval based on BoW model, and further proposed solutions from three aspects.Firstly, we analyzed the structure underlying the codebook, and proposed a novel classification method based on codebook structure. It took the density among codebook into account, which urged the neighbors of the local feature also to get highest affinity with each other. What is more, the locality and sparsity were well preserved during coding. As the experiments showed, the algorithm was able to adequately describe the semantics and efficiently improve the classification accuracy.Secondly, a new image coding approach based on regional features’coherence was presented by integrating features’ relationships among spatial neighbors and label cost. It made full use of the coherence and low-rank of regional features and realized the label optimization through a basis selection algorithm with locality label cost. Ex-perimental results indicated that the method could considerably improve the efficiency, while preserving accuracy.Finally, BoW model benefits to describe the high-level semantics, based on which we also proposed a new image retrieval approach that combined ant colony algorithm and probabilistic hypergraph. With the BoW representations, the method firstly en-hanced the affinity between the related images by ant colony algorithm. Then the task of image retrieval was converted to a hypergraph-based classification problem, and the retrieval results was obtained formally. Experiments validated that the method signifi-cantly outperformed the basis.