Large-scale Image Classification With Deep Clustering Network And Hierarchical Category Structure Learning

Image data has become an increasingly common media of information transmission on the Internet and mobile terminals.Although images contain a wealth of information,how to extract and utilize information in massive image data is still a problem in the field of computer vision.Image classification has a wide range of applications in smart security systems,automatic driving,and commodity search in e-commerce.This article focuses on large-scale image classification.Due to the large number of categories and large amount of image data in large-scale image classification problems.However,how to build a hierarchical structure between image classes based on visual attributes and classify them according to the hierarchical structure to improve classification efficiency.It is still a very challenging research topic.Large-scale image classification based on hierarchical learning has achieved some research progresses,but there is also a great space for improvement in the classification accuracy and the construction of hierarchical classification trees.This thesis proposes the following two major improvements for these issues:First,we propose a method for using Split Bregman algorithm to optimize joint dictionary learning,then use it for hierarchical large-scale image classification.In this paper,images are first grouped into multiple groups based on visual similarity.In each group,the joint dictionary learning algorithm is optimized by Split Bregman.After training,the group dictionary and specific category dictionary is obtained,group dictionary is used for group classification,and specific category dictionary is used for category classification.Experimental results show that using the joint dictionary algorithm optimized by Split Bregman improves the convergence,the classification accuracy is also improved.Second,a new deep auto-encoder subspace clustering network(DASC)is proposed.The DASC network structure is based on the deep auto-encoder network,and a self-expressive layer is added on the deep auto-encoder network.The parameters of the self-representation layer are the self-representation matrix parameters in the subspace clustering.The comparison experiment was conducted on the extended YaleB and COIL20.The experimental results show that the clustering accuracy of the DASC network is superior to other methods,and the category features of large-scale images are used as the input of the DASC network to cluster to build a hierarchical classification tree.Experiments were performed on ILSVRC2010 and Caltech256.The experimental results show that the hierarchical classification tree constructed based on the DASC network improved the classification accuracy.