Research On Context-integrated Multi-view IB Image Clustering Algorithm

With the rapid development of science and technology,the life of people has begun to step into the era of intelligent information.Image clustering is one of the key techniques in the field of Computer Vision,which plays a pivotal role in image retrieval.With the increasing complexity of the themes and scenes of image datasets,how to use the information in the images more effectively to improve the performance of image clustering has become one of the most important research direction in this field.For mitigating the effect of appearance variation on image,many scholars attempt to combine different views of images for clustering,i.e.,multi-view image clustering,where each view can be expressed as an image feature.In recent years,many classical and practical multi-view image clustering algorithms have been proposed.However,most current algorithms only consider the content information in the image or the context information across images,without integrating the advantages of both,which will inevitably have an adverse effect on the results of multi-view clustering.To solve this problem,this thesis proposes a novel Context-integrated Auto-weighted Multi-view Information Bottleneck(CAMIB)algorithm for image clustering with automatic weighted fusion of content and context.The CAMIB algorithm can cluster images using both content information and context information in the image data.The “content” describes the inner information of each image,such as texture,color and other appearance features;the “context” describes the close relationships among images in each view,such as the similarity among images.It also introduces the maximum entropy mechanism to automatically learn the weight of each view,so that the importance of different views can be integrated for effective clustering.In addition,in order to remove the extra weight regularization parameter in CAMIB algorithm,this thesis further proposes a Parameter-free Context-integrated Auto-weighted Multi-view Information Bottleneck(PCAMIB)algorithm.The algorithm regards image segmentation as a process of data compression,and by maximizing the preservation of the context and content information contained in the image when compressing the input images,the above problem can be formulated as an information loss function.Finally,a new "extract-merge" alternating iteration method is used to optimize the objective functions of the two proposed algorithms.Experimental results on five multi-view image datasets show that the CAMIB algorithm and the PCAMIB algorithm have higher stability and better clustering performance compared to several current state-of-the-art image clustering algorithms,and verifies the advantages of integrating context into multi-view image clustering.