Research And Application Of Graph Convolution Algorithm Based On Superpixel

With the strong feature extraction ability of deep learning of graph,more and more researchers have been looking into this field.Because graph is a kind of unstructured data,it can better describe the world and complete all tasks.In the field of computer vision,deep learning of graphics can accomplish image recognition,image segmentation and other tasks.These tasks can be used as the first stage feature extraction and image embedding by using super-pixel algorithm as image preprocessing method.At this time,image data is unstructured data representation.It is difficult to extract features by using standard convolution method,but it is easy to aggregate information and update nodes by using graph convolution method.At the same time,because a super-pixel aggregate dozens or even hundreds of pixels,it can achieve the information aggregation and redundant information removal,so that in some tasks,the difficulty of feature extraction can be reduced and the performance of the model can be improved.After introducing the common graph neural network and graph convolution algorithm,this thesis focuses on two tasks: image recognition and hyperspectral image classification.Firstly,a super-pixel graph convolutional neural network is proposed to realize image recognition.This is a challenging direction.In previous studies,most of the previous researches only use grayscale mnist-75 data set to complete image recognition task.It neither uses multiple data sets to verify,nor discusses the influence of super-pixel structure on the extraction of convolutional features of graphs.At the same time,few neural networks are designed to accomplish the task of super-pixel image classification.Based on this,this thesis uses the residual structure in deep learning,and combines the information of each layer with Concat operation to use the graph information of different granularity to explore the correlation between different layers and layers.At the same time,the loss function of improving the distance and difference between classes and the compactness within classes is used to monitor the whole network feature learning process.The feasibility of the method in image recognition is verified on multiple data sets,and the results of spatial domain method and spectrum domain method are compared.In addition,the thesis also explores the influence of the number of super pixels on the feature extraction of the graph convolution method.In the experiment,different number of super pixels are used on three common public data sets,and the effects of directed and undirected graphs on the neural network of graphs are also tested.Secondly,the thesis proposes a hyperspectral image classification method based on multiscale feature fusion,which combines the standard convolution method and the graph convolution method,and combines the super-pixel and pixel level features.Hyperspectral image classification is a point of every location in the classification data.Like the image segmentation in computer vision,but the convolution based model is designed for structured data.They often ignore some inherent relationship between adjacent land cover.Although the model based on plot volume can combine with the super-pixel to obtain the relationship between adjacent blocks,there is a lack of Local connection of adjacent pixel points.The method proposed in this thesis uses two branches to accomplish tasks.One is the convolutional neural network branch based on standard convolution,which uses space spectrum convolution as convolution module;the other is the graph neural network branch based on graph convolution,which uses residual module and concat operation used in the previous article.The two branches use the image information based on pixel points and the graph structure information based on the super-pixel block respectively.At last,the feature fusion of the two branches is completed,which can make the network learn multi-scale image information.The experiment is completed on multiple data sets,and the method using graph neural network only and convolutional neural network is compared.Finally,the effectiveness and robustness of the proposed method are obtained.