Research On The Supplement Of Image Annotation In The Social Media Environment

In recent years, with the rapid development of the Internet technology, there’sa rapid growth about the image, video and other multimedia information data on thenetwork. Gradually it became a major problem that how to manage, index, andretrieval these vast amounts of multimedia data effectively and accurately in thefield of image retrieval. One of the effective ways to solve this problem is toannotate these images, establish a mapping relationship between image visualfeatures and semantics features, and then the tags can be used to manage andretrieve images. So the technology of image automatic annotation which isbecoming a hot topic of researches at home and abroad, the research of it becomingmore widely and deeply.The main idea of this paper is: Based on the precondition of that the image hasbeen tagged, most images still can’t be described completely. In the light of thisproblem, this dissertation presents a method to supplement semantic imagedescriptions by associating a group of property tags with each existing tag. Thesetag properties can enhance the descriptive ability and the effective in tag-basedimage search of the existing tags. The main work and contributions are as follow:(1) Made an optimization to the feature selection and extracting of images onthe basis of the previous research work. Three different features are proposed inthis paper, and the information of image regions also are considered, theseeigenvector are quantified finally. The experimental results also demonstrate thatthis method reduces time complexity and space complexity.(2) This paper presents a method to calculate the similarity between region andtags, which combine the Near-far neighbors voting based method and the DiverseDensity Algorithm. The similarity measuring is the one key issue in this paper, weimproved the calculating method of Score, and the k-farthest image as a negativebag is considered, then increase the weighting of the k-nearest images which areconsidered as positive bag in the calculation of Score. Results show that，afterconsider the k-farthest images as the negative bag, and adjust the weight of positivebag and negative bag, the value of the score have greater variance, improve thediscrimination of different annotation, helps us to judge the correlation dimensionconveniently. (3) Aim at improving the accuracy and reliability of the supplemented propertytags, the prior knowledge of existing tags were used to construct a priority table.After adding the property tags, this table were used to improve correctness of theadded tags in the supplement process of color、texture and shape tags. This methodis able to consider the prior knowledge of existing tags, which is helpful to improvethe accuracy of the added tags.