Using D-S Evidence Theory To Incorporate Multiple SVMs For Image Annotation And Retrieval

Content-based image retrieval refers to find the images which are same or similar with the given query image from a large image database according to the images content. However, the―semantic gap‖of low-level features and high-level semantics of images is key issue for image retrieval, in order to improve the performance of image retrieval system, in this paper, a novel image retrieval system is proposed, which integrates two sets of support vector machines (SVMs), namely global-based SVM and Co-SVM, for image retrieval. The global-based SVM is trained based on the concatenation of global color and texture features, whose output is modified by using Evidence Theory to obtain the ?nal classified results. The Co-SVM is learned respectively in color and texture feature of labeled images with two different kernel functions according to the distinct statistical properties of the features, whose outputs are incorporated by D-S Evidence Theory. Then we can get confidence degrees of unlabeled samples according to the results of fusing and select disagreement and agreement samples from unlabeled data. We also propose a new active feedback algorithm utilizing more positive samples to reduce the cost of labeling and improve the precision of classifiers. Given an input image, the goal of automatic image annotation is to assign a few relevant text keywords to the image that re?ect its visual content. We use the current popular methods of image retrieval and treat the problem of annotation as image retrieval; we get help from the comprehension of users according to the two feedback frameworks. The first feedback framework is modified the probability of output for the multi-class SVMs by using the D-S evidence theory, when retrieval the images which are similar with the query image, users can improve the efficiency and accuracy of image search by relevant feedback. The second feedback is obtained visual information of images through the first feedback framework; we take the measure of assurance which is modified by evidence theory as the weight of visual information of images. Then we use high-efficiency marker transfer strategy, in order to transfer some keywords of similar images to the query image, so we can get the keywords which re?ect its visual content, at the same time, the users can label the keywords based on their comprehension of the image's visual content by feedback of keywords, in order to improve the accuracy of image annotation.According to the algorithm that we proposed, we design and develop an image retrieval and annotation system based on Using D-S Evidence Theory to incorporate multiple SVMs. Experiments show our algorithm can reach a good result in a much less rounds of feedback compared to other algorithms.