Research And Application Of Image Classification Algorithm Based On Projected Entangled Pair States Tensor Network

With the rapid development of the Internet,the related multimedia data is also increasing exponentially.These multimedia data contain a lot of valuable image,and the processing method for these image data-computer vision,has become a very frontier research topic in artificial intelligence.Image classification,as the basic task of computer vision,is the basis of many more complex tasks.For image classification tasks,the current main achievements are concentrated in the direction of neural networks.The main method is to first use convolutional layers to abstract features from images layer by layer,and then use classifiers to perform classification tasks on features.The emergence of deep convolutional neural networks has further promoted the application of computer vision in various directions,which provides a solid foundation for research in other fields.For example,the method based on Transformer has achieved excellent results in vision.The tensor network method based on quantum physics has also gradually entered the field of computer vision,and has developed a new algorithm by virtue of its own characteristics.The tensor network model that was first applied to image classification was Matrix Product States(MPS),and its application in the Fashion-MNIST dataset proved the feasibility of tensor network as a machine learning model.Afterwards,a variety of MPSbased tensor network models emerged,but most of these models need to expand the image into a one-dimensional vector as input,which will lead to the loss of image structure information.At this time,the Projected Entangled Pair States(PEPS)tensor network with a similar structure to the picture is introduced into the image classification task.Since PEPS has a two-dimensional structure,it can naturally obtain structural information from pictures,and it has achieved better performance than tensor networks such as MPS in MNIST and Fashion-MNIST datasets.However,compared with the mature neural network,the classification model based on tensor network still has a big gap in performance.In this thesis,based on the PEPS tensor network,combined with the idea of network layering in deep learning and locally disordered tensor networks,a Multi-Layered Projected Entangled Pair States(MLPEPS)tensor network is proposed for image classification task.The input image information is abstracted layer by layer through cascaded PEPS layers,and MLPEPS constitutes a brand new tensor network model.MLPEPS utilizes the two-dimensional network structure of PEPS to obtain the picture structure information,and at the same time,integrates features through higher-level PEPS blocks.At the same time,both the number of PEPS layers and the structure size of PEPS blocks can be dynamically adjusted as hyperparameters,which makes MLPEPS more scalable.MLPEPS obtained better test set accuracy than other tensor networks on the Fashion-MNIST dataset,and surpassed many classic machine learning methods.Afterwards,a comparative experiment was carried out on a more realistic COVID-19 data set,and the test set accuracy of MLPEPS reached 91.63%,which is close to the performance of Goog Le Net.In order to further tap the potential of PEPS in image classification,based on the characteristics of the fully connected layer in the convolutional neural network,this thesis proposes to use PEPS as the tail classifier of the convolutional neural network,so as to optimize its performance.Experiments on the COVID-19 dataset show that the PEPSoptimized neural network can achieve better performance,and in some models,the compression of model parameters has also been achieved.By exploring the application of PEPS in image classification algorithms,this thesis further explores the potential of tensor network in image classification,and at the same time provides more application scenarios for tensor network in image classification.