Research On Web Image Annotation Based On Context Information

As the information technology has developped rapidly,the Internet is playing an increasingly important role in human's daily life.At the same time,there are a great number of information data on the Internet.With the development of multimedia technology,the data forms for description and transmission tend to diversity.In addition to the text information,the other forms of data,such as image,video,become the main data forms on the Internet.The image has the characteristics of vivid and understandable,dilivers meaningful information to users.Not only as an independent data form,the image can also be complement with other forms of data.The characteristics make image the indispensable on the Web.As a typical form,Web image is embeded in the Web.The research has found that Web image has rich textual information around exclude the visual feature,namely the context information of Web image.The context information of Web image always reveal the topic content,and it is great significant for Web image automatic annotation.The thesis aims to research on the Web image automatic annotation,making full use of both the native features and the rich context information of the Web image to tag them.The research on extraction and merging scheme of Web image features and the framework of annotation could help to promote the performance of annotation,and establish a good description mechanism for Web image,so that the Internet users can retrieve tons of images more accurately.Moreover,the research is of great significance to the research of multimedia.The main and innovative works are as follows:(1)The weighted solution for context information based on AHP is proposed.At first,the types of context information for Web image are discussed.Considering the different semantic contribution depended on their own types of context information,the weighted solution for context information based on AHP is proposed,to estimate the weight of different types of Web image context information.(2)The above weighted solution is employed during the textual features extraction of context information.To get a more accurate result,the weighted coefficient is imported to the TF-IDF algorithm,which is the popular method used to natural language processing including the Chinese processing.As a result,the set of keywords,which embodies the context information,is prepared for further merging in the following process.(3)The topic model based on LDA is used to extract the native semantic features of Web image in the research.The thesis selects the SIFT feature as the low-level semantic feature,and makes use of the clustering technology to generate the Bag of Visual Words;Therefore,with the help of this Bag of Words,the topic model based on LDA generate the set of keywords,which embodies the graphic features.(4)The merging scheme for context features and graphic semantic features is proposed,which is based on PageRank algorithm.In order to annonate Web image more quickly and efficiently,the thesis intents to merge textual and visual information of Web image,both of which are generated by the above two steps.As a famous ranking algorithm,PageRank algorithm takes the number and quality of links into account to determine the importance of a certain website.The thesis converts the merging problem into a ranking problem with more than one factor,and the PageRank is employed to rank the keywords which can annonate a Web image.(5)Finaly,the thesis figures out the framework of Web image automatic annotation.The general framework of automatic image annotation is studied,and the thesis specifies the framework by importing the uniqueness of the Web image automatic annotation,and the prototype system of Web image automatic annotation has been implemented.In this thesis,for the algorithm proposed in the above-described works,experiments had been implemented on the data set.And the results verify the effectiveness of the framework of Web image automatic annotation.