Research On Convolution Neural Network Based Hyperspectral Imagery Super-resolution Restoration

Hyperspectral image is a three-dimensional data structure that is different from ordinary two-dimensional image.This kind of image can simultaneously acquire the spatial dimension and spectral dimension information of the target.When imaging,in addition to the visible light band,it can simultaneously image in multiple spectral ranges,which can describe the target information in more detail.However,the spatial resolution of such images is usually lower than that of ordinary two-dimensional images,which leads to great limitations in applications such as identification or detection.Therefore,restoration of high-resolution images by super-resolution restoration techniques of signal processing has become an important way to improve the spatial resolution of images.However,the existing algorithms usually only convolve and reconstruct the spatial dimension high-frequency information of the two-dimensional ordinary image,failing to utilize a large amount of useful information contained in the image spectrum dimension,and the restoration effect is not good.In this paper,based on the data characteristics of hyperspectral imagery,a hyperspectral image superresolution restoration algorithm based on three-dimensional convolution is studied.The main research contents include:According to the characteristics of hyperspectral image space and spectral dimension correlation,a three-dimensional convolution kernel for super-resolution restoration applications is designed,and the spatial and spectral dimension features of hyperspectral image are extracted to compensate for the lack of spectral dimension feature utilization of two-dimensional convolution kernel.Furthermore,the effects of various parameters of the convolutional neural network on the feature extraction of the three-dimensional convolution kernel are analyzed by simulation experiments.The convolution layer number,convolution kernel size and activation function are optimized for high-performance super-resolution restoration.The research of the algorithm laid the foundation.Based on the realized three-dimensional convolution kernel,a hyperspectral image super-resolution restoration algorithm based on three-dimensional residual dense network(3D-RDN)is designed and implemented.The RDN network is used to make full use of the advantages of the hierarchical features of deep convolutional networks.The multi-dimensional space and spectral dimension features of hyperspectral images are simultaneously extracted by three-dimensional convolution kernels,and the corresponding three-dimensional sub-pixel convolution links are designed to achieve high Super resolution recovery for performance.The experimental results show that the proposed method has higher utilization rate for the hierarchical features extracted by all convolutional layers,and the results are significantly better than the existing superresolution restoration algorithms.Furthermore,in the 3D-RDN algorithm,the hyperspectral image spectral dimension information is directly introduced due to the three-dimensional convolution kernel,resulting in too many parameters,long running time,and high GPU memory requirements for the training model.A super-resolution restoration algorithm based on Virtual-3D convolution kernel.According to the tensor decomposition theory,the threedimensional convolution kernel is decomposed into a combination of low-dimensional two-dimensional convolution kernel and one-dimensional convolution kernel,and the best combination of performance is determined through simulation experiments to construct the Virtual-3D convolution kernel.Based on this,the improved 3D-RDN algorithm is further optimized.Experiments show that the algorithm not only reduces the overall computational load of the network,but also improves the super-resolution recovery performance.