Integrating Visual and Textual Features for Web Image Clustering

With the explosive growth of Web and tremendous development of digital image processing technologies, the applications of Web image have attracted much attention, such as the Web image retrieval. Since the Web images are often with some related text tags, making use of both visual and textual features of Web image will help improving the accuracy of the Web image clustering. Researches show that Web image clustering methods, such as graph partitioning models and hyper graph partitioning models, didn´t make use the relations between texts and image simultaneously. In this paper, we explore to take both visual and textual features into account for Web image clustering by building a graph model and develop a novel iterative clustering method. With K clusters initialized, we calculate the occurrence frequency of each visual/textual feature over the j-th cluster (j = 1, 2,⋯, K), which is used to measure the significance of the feature for the j-th cluster. Then the likelihood of each image, which belongs to the j-th cluster, can be determined accordingly. Furthermore, a mixture model is built for the predicted feature linked to each image and the EM algorithm is adopted to get K component parameters which describe posterior probabilities of all clusters for each image. Then two K-dimensional vectors consisting of component parameters will be used to describe the image and adjust the cluster index of it. Several experiments have been performed with MIR-Flickr25K and IAPR TC-12 Benchmark datasets and the performance of the proposed Web image clustering algorithm is superior to that of the compared algorithm.