Image Retrieval Method Collaborative Tagging Research

The information sharing and on-line collaborative in Web2.0let every user not only can get all kinds of information from the Internet, but also can be free to release all sorts of information or labeling for it. A common application is labeling the articles on BBS, it is convenient for classification and retrieval. This method that labeled by users is called the cooperative labeling. A large number of users’labels on internet provides important information for rerieval. Now this collaborative annotation always use words or phrases, it is enough for text information, such as the blog article or BBS posts, to achieve the purpose of retrieval by the labels, but the labels can’t fully accurately reflects the content of the data and structured information for some high-dimensional data such as Internet image or geographic information. Therefore, it is necessary for images on Internet to establish a new retrieval model, Connecting the labels by people according to understand the images with image visual content and set up images topic network so as to improve the image theme image retrieval accuracy.The image of the topic network will help user interactive image retrieval. At the same time, the topic network contains semantic information of images and visual information of images, so it can greatly improve precision and recall ratio of the image retrieval. In this paper, we design a new image retrieval model, users do retrieval through the image topic network that contains the similarity content between topics. The image extraction is through the filter rubbish tag and evaluation function of interested labels, The similarity of the topics use the method included the semantic similarity content of topics and visual similarity content of topics.The semantic similarity content of the topics is got by the WordNet semantic similarity of the labels as topics and the co-occurrence frequency of labels as topics; the visual similarity content of topics is got by the mixed kernel feature. First, extracting image color histogram, Gabor texture feature of the image as a global visual features.second, extracting based on local color histogram based on overlapping block as the partial visual features, get the mean of each feature in the same dimension under the same image topic as the visual features of the topic. Third, the three visual feature of topics is mapped to kernel space, the last, three kernels are mixed to form a mixed kernel feature that describes visual similarity content of topics. In order to be more effective to display images topic network or image retrieval results, to show more image information at limited plane, this paper use hyperbolic visual method to design a image retrieval interface, the method visually displays the topic network in one interface. Users can drive through the topic network by moving a mouse for interactive browsing retrieval when they use the topic network. At the same time, the user can also input search by inputing a keyword and retrieval results will be displayed visually. The experiments proved that the validity of this paper’s method.