Research On Image Super-Resolution Reconstruction Algorithms Based On Multi-Level Degradation

With the development of science and technology,low-resolution(LR)images are difficult to meet the needs of many fields,and the demand for high-resolution(HR)images is increasing day by day.Because the improvement cost of imaging equipment is too high and the imaging technology is not up to standard,it is difficult to obtain HR images by upgrading the imaging equipment.Super-resolution(SR)image reconstruction is a technology to obtain HR images through algorithm design at the software level.Because of its flexibility,economy,and applicability,SR has good application value in both daily life and production activities.With the rapid development of deep learning,the SR image algorithm based on deep learning has an unprecedented development opportunity.Aiming at the problem of SR reconstruction of degraded images by bicubic interpolation and the problem of SR reconstruction of real degraded images,this dissertation proposes the corresponding image SR reconstruction algorithms.The main research contents are as follows:(1)Aiming at the problem that the deep neural network image SR reconstruction algorithm with attentional mechanism is difficult to extract the multi-level high frequency features with different difficulties,an attentional hierarchical network image SR reconstruction algorithm is proposed.Firstly,an improved dense block is designed to fully extract low-frequency features.Then,an attention block is designed to separate the easy-to-extract components and the difficult-to-extract components from high frequency features,and a dense residual module is further designed to extract the deep features of the difficult-to-extract components.Finally,the attention hierarchical feature extraction module is designed,which separates and extracts high-frequency features with different difficulty by layering,so as to obtain multi-level high-frequency features.A large number of experimental results on Set5,Set14,BSD100 and Urban100 test sets show that compared with mainstream algorithms,the proposed algorithm has improved PSNR and SSIM,and achieved better SR reconstruction performance.(2)Aiming at the problem that the image SR reconstruction algorithm based on deep neural network is easy to lose feature information during feature extraction,resulting in the lack of texture and edge details in the reconstructed SR image,a U-shaped network image SR reconstruction algorithm with multi-level information compensation is proposed.Firstly,a U-shaped network for image SR reconstruction is designed.The network performs multi-level feature extraction and channel compression for the input features through the down-channel branch,fuses the compressed features through the bottom module and extracts the related features of different channels,and performs multi-level feature extraction and channel recovery for the compressed related features through the up-channel branch.Then,a multi-level information compensation model is designed,which compensates the lost information in the channel compression process of U-shaped network and the information which is difficult to recover in the channel recovery process.Finally,the proposed algorithm and mainstream algorithms are tested and analyzed on Set5,Set14,BSD100 and Urban100 test sets.The experimental results show that the proposed algorithm achieves a great improvement in PSNR/SSIM and visual effect compared with the mainstream algorithms.(3)Aiming at the problem that the convolution layer with different convolution kernels or multi-branch deep neural network image SR reconstruction algorithm ignores the correlation of multi-scale context information and high-frequency components of LR images,a fusion attention network image SR reconstruction algorithm based on dilated convolution is proposed.Firstly,a dilated convolution attention module is designed,which captures multi-scale context information by locking multiple areas of receptive fields with different sizes in LR images.Then a multi-level feature attention block is designed to further focus on the high-frequency components of multi-scale context information and extract more high-frequency features.The proposed algorithm is tested on four benchmark test sets,and the results show that the proposed algorithm has good effects in quantitative evaluation index and visual effect.(4)Aiming at the problem that the image SR reconstruction algorithm based on convolutional neural network is difficult to extract the multi-level feature information in the LR image,which leads to the lack of rich details in the reconstructed SR image,a multi-level continuous encoding and decoding image SR reconstruction algorithm is proposed.Firstly,this algorithm extracts shallow features from LR images through initial convolution layer.Then,different levels of image features in the LR image are acquired through a plurality of end-to-end connected multi-level continuous encoding and decoding attention residual modules,and different weights are generated for these features according to different extraction difficulties,and the image features at different levels are recalibrated,so that rich detailed features in the image can be acquired;Finally,the extracted rich detail features and shallow features are reconstructed into HR images by the upsampling module and the reconstruction convolution layer.Through the comparative test on four benchmark test sets,the results show that the SR image reconstructed by the proposed algorithm is superior to the image SR reconstructed by mainstream algorithms not only in objective evaluation index but also in visual effect.(5)The image SR reconstruction algorithm based on deep learning improves the image SR reconstruction performance by increasing the depth of neural network,which leads to the problems of network model complexity and excessive parameters.To solve these problems,an inverted N-type lightweight network image SR reconstruction algorithm based on back projection is proposed.Firstly,the initial convolution block is used to extract the shallow features of LR images.Then,the deep features of the LR image are extracted by two rounds of gradual model compression and two rounds of gradual model recovery of the inverted N-type network based on back projection.Then,the extracted deep features and shallow features are combined by global residual learning,and amplified to the desired output reconstructed SR image size by the upsampling module.Finally,the reconstruction module is used to reconstruct the SR image.The experimental results on four benchmark sets show that the proposed algorithm has a lighter network structure than mainstream algorithms,and the reconstructed SR image not only has higher PSNR and SSIM,but also has better visual effect.(6)Aiming at the problem that the image blind SR reconstruction algorithm that only uses blur kernel as the prior knowledge of SR reconstruction process to cause a large estimation error of blur kernel,a blind SR reconstruction algorithm based on parallel alternating iterative optimization is proposed.The algorithm uses the dynamically corrected blur kernel and the dynamically extracted LR image features as the prior knowledge of the SR reconstruction process and the blur kernel correction process to iteratively correct the blur kernel.At the same time,a blur kernel updating module is designed,which takes the extracted LR image features as the prior knowledge of the SR reconstruction process.An enhanced spatial feature transform residual block is designed,which takes the corrected blur kernel as the prior knowledge of the blur kernel correction process.A large number of experiments on two synthetic data sets and a real scene LR image show that the proposed algorithm is superior to the current mainstream blind SR reconstruction algorithms,and can reconstruct satisfactory SR images.