Super-resolution Reconstruction Of Remote Sensing Images Based On Sparse Representation And Dictionary Learning

Remote sensing image is an important means to obtain ground information,which has been widely used in forest,grassland cover detection,wetland resources monitoring and other aspects.The super-resolution reconstruction technology of remote sensing image is a technology to obtain high-resolution image at low cost by using software processing method to break through the limitation of imaging system under the existing remote sensing imaging conditions.In this thesis,on the basis of sparse representation and convolutional sparse coding,super-resolution reconstruction is studied by using multi-channel fusion and online dictionary learning theory,and two super-resolution reconstruction algorithms are proposed.Based on the four-channel sparse representation super-resolution reconstruction algorithm,the images are sequentially turned 90 degrees,and each channel image is reconstructed using an improved sparse representation algorithm.Firstly,an adaptive median filter is used to preprocess the learning samples.The feature extraction of the processed samples is carried out to get the sample blocks needed for training,and then the sample blocks are used to train to get the high and low resolution dictionaries.Then,the high-resolution image is reconstructed by the combination of sparse representation coefficient and high-low resolution dictionary.Finally,the reconstructed four-channel image is weighted and fused to get the final high-resolution image.Compared with other algorithms,the objective value of PSNR is increased by 1d B and SSIM is increased by about 0.001 and the FSIM is improved by about 0.02.Based on the online convolution dictionary learning algorithm,the image to be reconstructed is divided into high and low frequency parts,and the low frequency part is reconstructed based on the four-channel algorithm.the high-frequency part uses the online convolution dictionary learning method to train the high-resolution online convolution dictionary and the low-resolution online convolution dictionary.Then,the mapping function between high and low resolution online convolution dictionaries is obtained,and the low resolution images are reconstructed with the online convolution dictionaries.Finally,the reconstructed high resolution images are merged to get the finalhigh resolution images.Compared with other algorithms,the objective value of PSNR is increased by 1d B,and the SIM is increased by about 0.01.The FSIM is improved by about 0.001.Some simulation studies are carried out on the two algorithms proposed in this thesis,and it is verified that the image reconstructed by the algorithm proposed in this thesis is clearer subjectively,and the objective value is also improved compared with the images obtained by other algorithms.