Research On Super-resolution Method Of Hyperspectral Image Based On Fusion Convolution Neural Network

Hyperspectral image super-resolution technology is one if the very important hyperspectral remote sensing technologies.The traditional super-resolution methods use the analytical methods such as spectral end element extraction,spectral unmixing,and sub-pixel positioning to obtain high-resolution hyperspectral remote sensing images.Due to the shortcomings of analytical methods in flexibility and super-resolution reconstruction quality,searching for a super-resolution method with high applicability and high reconstruction quality has become a new research focus in this field.In this paper,the super-resolution method of hyperspectral remote sensing image based on fusion convolutional neural network is studied deeply.The main contents are as follows:(1)In order to solve the problem of insufficient cognition of response characteristic of convolution operation in neural network,the reasons for the response of convolution layer in neural network changing with various parameters are obtained based on the theoretical analysis of the response characteristics of convolution operation.In combination with the convolution network experiment,the response characteristic change of the convolution process is explained.(2)To solve the problem that the existing methods can’t make full use of the spatial information of different frequencies of hyperspectral data,and the problem that the existing high quality hyperspectral remote sensing image data is insufficient as well,a multiscale convolution super resolution network based on wavelet transform is proposed.Each frequency band of hyperspectral remote sensing images is independently processed,using twodimensional wavelet packet.The wavelet coefficients of different frequencies are obtained,features of which are then extracted by a multi-scale convolution branch.Then high-resolution wavelet coefficients are obtained through sub pixel convolution and detail reconstruction layers,outputs of which finally consist the super resolution image by wavelet coefficient reconstruction.In the process of back propagation,the network parameters are optimized by calculating the loss value between the predicted wavelet coefficients and the real wavelet coefficients.Experimental results show that the super resolution images obtained by this method have better peak signal-to-noise ratio and structural similarity,achieving the goal of making full use of spatial information of different frequency elements of hyperspectral data.(3)A multiscale convolution super resolution network based on wavelet transform is proposed to solve the problem of spectral distortion in the previous methods.The network utilizes the three-dimensional wavelet transform to obtain the wavelet coefficients of different frequency elements,features of which are extracted by the multi-scale three-dimensional convolution branches.Then the high-resolution wavelet coefficients are obtained by the threedimensional deconvolution and detail reconstruction layers,outputs of which finally consist the super resolution image by three-dimensional wavelet reconstruction.In the back propagation of the network,the network parameters are optimized by calculating the loss value of the predicted wavelet coefficients and the real hyperspectral remote sensing image wavelet coefficients.The experimental results show that the super resolution images obtained by this method not only have high super resolution quality,but also effectively suppress spectral distortion.