Research And Application Of Multi Paramter Graph Signal Processing

Graph signal processing with its application in image processing has been a hot spot in domestic and foreign research recently.Based on graph signal processing theory and graph deep learning methods,this thesis aims at the challenges of image interpretation and information security in digital image processing applications,including the non-European structured representation of data,deep feature extraction and analysis,etc.Firstly,for the application of graph signal processing(GSP)theory in image encryption,extend the graph Fourier transform to the angle graph Fourier transform(AGFT)so that it has both the multi-parameter characteristics of the angle transform and the structure-related characteristics of the graph transform.A digital color image encryption algorithm based on the Multi-Parameter Angle Graph Fourier transform is proposed,considering both the pixel intensity and internal structure,and combines with DNA coding technology,which effectively realizes the secure transmission of image data and achieves better performance than existing algorithms.Then for the application of remote sensing image classification,a multi-scale graph convolutional neural network is applied to feature extraction of hyperspectral data at different spatial scales.A two-branch CNN method is used to obtain the spatial-spectral feature of multi-source data;then a random walk layer is used to combine prior distribution,pixel correlation,and training samples for classification.A multi-source collaborative classification method is proposed based on a multi-scale graph convolutional neural network(MGCN),which combined the features of the hyperspectral image and Li DAR images to effectively achieve the fine classification of urban land-covers.This thesis has carried out in-depth research on the graph signal processing theory and its application to image encryption and remote sensing image classification.Simulation experiments further confirmed the effectiveness.