Multivariate IB Clustering Algorithm Based On Image Visual Context

Because people are living in an intelligent and technological world,mobile camera devices are becoming more and more popular in our daily life.With the increase of image capture tools,the way of obtaining images is more simple and convenient,and the amount of image data is much larger.In recent years,more image visual features have been used to describe images and how to improve the clustering quality of massive image by effectively using the connections among multiple features has been the popular and difficult problem in current research on image clustering.In recent years,algorithms for multiple image features have been proposed,which can be divided into two categories: ensemble clustering methods and multi-view clustering methods.Ensemble clustering methods first obtain multiple basic clusterings,then use a fusion function to re-clustering the basic clusterings in order to get a better result.Ensemble clustering methods are over dependent on basic clusterings and neglect the original feature of image data,so the performance of ensemble clustering methods is limited to the original basic clusterings.Multi-view clusterings can deal with multiple visual features of image data,as multiple views,and obtain a partition better than the clustering result of single visual feature by associating multiple visual features of image data.The multi-view clusterings may have the problem of dimensionality curse when processing image visual feature with high dimension.To solve the above problems,a novel image clustering algorithm called VC-IB(Visual Contextual Information Bottleneck)is proposed.VC-IB regards data analysis as the procedure of data compression and randomly chooses one of the visual features as the main information and treats the clusterings obtained by the remaining visual features as the visual contexts in the process of data compression.VC-IB can mine the latent patterns in the image data by considering both the main information and visual contexts of the image data.Therefore,VC-IB inherits the advantages of both ensemble clustering methods and multi-view clusterings.In this algorithm,two Bayesian networks are constructed to describe the connections between data compression and information preservation,a sequential solution called “draw and merge” is utilized to optimize the objective function,and mutual information is used to measure the information between the main information and the visual contextual information of the image.The experiments on five image datasets show that the proposed algorithm VC-IB can effectively process the visual contexts of image data and consistently and significantly outperform other state-of-the-art clustering methods.