Collaborative Image Retrieval And Annotation

With the widespread use of the Internet and the popularity of digital products,digital image is increasingly important in people’s lives. In order to find the images thatusers need from the massive amounts of digital images quickly and accurately, imageretrieval technology was proposed. Now, the main research focus of image retrievaltechnology are content-based image retrieval and semantic-based image retrieval.Semantic-based image retrieval is based on image semantic. The automatic imageannotation is the main method to obtain semantic for image. Image retrieval relevancefeedback logs contain many potential information, these information reflect users’judgement and evaluation for image retrieval result and similarity. These informationcan help to improve accuracy of image retrieval and quality of image annotation. Thepaper focus on the feedback logs for image retrieval and image annotation. The paper’smain contents are as follows:First, the paper present a method to collect, store and aggregate feedback logs.Collect feedback log when users submit relevance feedback, use index table to store thefeedback logs, and aggregate feedback logs to a matrix. The matrix can be updatedeasily, and can get correlation of images quickly. The index table can save store space,and convenient for analysis logs when using other method.Then, make use of the feedback logs in image retrieval, and propose a collaborativeimage retrieval algorithm based on feedback logs for image retrieval. Relevancefeedback is the main method to solve the "semantic gap". To solve the small sample sizeproblem of relevance feedback, many researchers make use of support vector machine(SVM) in relevance feedback. But, in general, the sample size of relevance feedback istoo small to get a good effort for image retrieval. In this paper, to solve the small samplesize problem of relevance feedback better, a way to expand the training sample based onfeedback logs is proposed. Considering the different importance of training samples,weight support vector machine (WSVM) is used to train the classifier. Based on aboves,the accuracy of retrieval is greatly improved.Next, make use of the feedback logs in image annotation, and present a algorithmthat combining feedback logs and hybrid probabilistic model (HPM) for imageannotation. HPM use image’s low-level features and tag word thesaurus to tag image,and achieve fairly good results. But HPM also ignore the useful information in feedback logs. The paper analyze the feedback logs and mining the relationship between the tagsin specific application in this paper. Then replace the relationship between the tags inthe initial HPM. Based on above work, the quality of automatic image annotation isimproved.At last, a framework which including image retrieval and image annotation is builtin the paper, and this framework can be named as a collaborative image retrieval andannotation system. This system is a simple prototype system and developed by usingMatlab. Content-based and simple tag-based image retrieval can be carried out on thissystem. Also, this system can update the tags of image according to the size of feedbacklogs.