Research On Image Super Resolution Reconstruction Based On Deep Learning

Image super-resolution(SR)is a classical low-dimensional computer vision task that aims to recover a degraded low-resolution(LR)image into a visually gratifying high-resolution(HR)image.The image SR problem is essentially an ill-posed problem without a unique solution,and it is difficult to solve this problem by mathematical expressions alone.In recent years,deep learning-based image SR methods have started to emerge.The current deep learning-based image SR methods deepen the network in order to obtain better reconstruction performance,but this approach makes the entire network structure too parametric and computationally intensive,which limits the application of these algorithms on mobile devices.In addition,most of the existing studies are based on ideal degradation models(e.g.,bicubic degradation),while the degradation models of images in real scenes are very complex and difficult to model.Therefore,the network trained with the degradation model of bicubic downsampling will have a relatively large performance drop in real scenes.In this paper,we conduct an in-depth study of the above problem,and our main work is as follows:(1)To address the high computational complexity of existing image SR reconstruction algorithms,a lightweight image SR reconstruction algorithm based on adaptive sparsity is proposed in this paper.A shallow feature extraction block is constructed by introducing a sparse self-attention mechanism,which is able to extract shallow features containing global information at the cost of less computation,laying a solid foundation for recovering highfrequency details of images.In order to take into account the inherent sparsity in image SR tasks,an improved sparse mask block is proposed that can skip the redundant computation of image flat areas during the forward inference of the network to save computational effort.A feature fusion block is then designed to fuse the outputs of the previous cascaded sparse mask blocks to provide features with more high frequency detail to the image reconstruction block.In addition,the entire SR network uses skip connections between blocks to ensure that the low-frequency details of the image are not lost.It is demonstrated that the proposed algorithm achieves better reconstruction results than other lightweight SR algorithms with less computational complexity on all public test sets.(2)This paper proposes an enhanced blind super-resolution reconstruction algorithm based on unsupervised degradation representation learning to address the phenomenon that existing image SR reconstruction algorithms cannot recover high-quality images under the degraded model of real-world scenes.The algorithm mainly consists of two branching networks.The first branch is a residual-based encoder network in the contrast learning framework,which can learn a high-dimensional abstract degradation representation vector from the input low-resolution image.And this degradation representation vector can provide the SR reconstruction network with information about the degradation model of the lowresolution image to be reconstructed.Another branch is the degradation-aware fusion SR network that can flexibly adapt to various degradation models using the degradation representation vector learned by the encoder.The backbone of the degradation-aware fusion SR network is the degradation-aware fusion block,whose core idea is to fuse the input LR image features and the degradation representation vector using three branches.In addition,a shallow feature extraction block is designed for the degradation-aware fusion SR network in order to extract features containing global information from the input low-resolution images.Extensive experimental results on synthetic datasets and real images show that the proposed algorithm is able to achieve a leading reconstruction performance over other blind SR algorithms for the blind SR task.