Research On Image Representation Via Multi-Graph Embedding

With the widespread popularity of Internet and the rapid development of information technology,the scale of image data are increasing.Therefore,how to obtain the most reasonable and valuable feature representation from massive image data has become a hot topic in the fields of image information technology and pattern recognition.Aiming at this problem,this thesis proposes a multi-graph embedding technique and applies it to manifold alignment and deep auto-encoders to obtain better feature representation capabilities.In addition,an image recognition system is designed and implemented based on the proposed algorithm.In this thesis,a novel manifold alignment algorithm based on multi-graph embedding technology is proposed.Inspired by the construction process of traditional graph embedding algorithms(LPP,NPE,SPP),we propose a multi-graph embedding framework which can fuse multiple graph embedding matrices constructed from different angles to obtain more comprehensive feature information in the manifold.Then,the framework is applied to the semi-supervised manifold alignment algorithm,so that the local structure information of each manifold during the alignment(multiple manifold joint projection)process is better preserved in order to obtain better alignment presentation.The experiment results on multiple datasets demonstrate superior performance of the proposed method in terms of accuracy and robustness.In this thesis,an auto-encoder ensemble model based on multi-graph embedding technology is proposed.First,based on the idea of ensemble and the diversity of activation functions,the traditional deep autoencoder is extended to a multi-encoder parallel integrated network structure to save more effective feature information and improve the generalization ability of the model.Furthermore,in order to maintain the local connectivity from the original image space to the feature subspace,which means the feature representation obtained from the encoder model still maintains the manifold structure of the original image data,we further applied multi-graph embedding technique in above ensemble model.The classification results of classic visual datasets such as MNIST and Cifar-10 show that compared with the traditional deep autoencoder model,the auto-encoder ensemble model can obtain more comprehensive feature information and maintain the local manifold structure of input data.An image recognition system is designed and implemented based on the proposed feature representation algorithm.It is composed of system management and image recognition.Several running tests show that the image recognition system constructs a friendly functional interface,achieves the required functions,and verifies the practicability of the proposed image feature expression algorithm.