Spatial Contextual Information Based Image Retrieval

With the development of multimedia technologies, image retrieval has beenwidely used in many aspects of social life. We propose two image retrieval methodscombined with the spatial context based on BoW model. One is compact visualphrases based image retrieval; another is the longest common visual substring basedimage retrieval. The paper is summarized as follows.Firstly, this paper reviews the history and present status of image retrievaltechnology, and introduces local features of images. Through in-depth analysis ofBoW, the paper argues that lack of contextual information is a key factor which isrestricting retrieval performance.Secondly, chapter III presents bi-gram visual phrases based image retrieval. Themethod constructs visual phrases in significant areas which are salient regions. In thismodel, the number of salient regions and the radius of the salient region areconsidered as key factors that affect the construction of visual phrases, and twofactors are analyzed by experiment. BoW, Saliency-BoW and Saliency-BVP arecompared in the experiment and the experimental results demonstrate Saliency-BVPis the best among them. This shows that visual phrase contains more information thana simple visual word. Hence, spatial contextual information for image description isvery important.Finally, Chapter IV presents an image retrieval method based on longest commonvisual substring (LCVS). This idea is that hidden pattern between the two imagesconstructs LCVS. It regards LCVS as the pattern of occurrence between two images.The method achieves the image retrieval by regarding the value of LCVS as similarityof two images. Then, this chapter also proposed a method for calculating thesimilarity of common string, called the method common substring weightedmaximum (CSWM). Experimental results show LCVS have a better performance thanBoW in public image dataset. We also compare CSWM and LCVS in the experimentsand found almost the same results. It further shows that the spatial contextualinformation can not be ignored.