Research On Lightweight Single Image Super-Resolution Algorithms

Super-Resolution(SR)is the reconstruction of a Low-Resolution(LR)image into a High-Resolution(HR)image.Image super resolution is also a hot and difficult problem in computer vision.A high-resolution image degenerates into a low-resolution image with different degradation directions in different environments,so a high-resolution image can have multiple low-resolution images.The problem is essentially a many-to-one relationship,so it is difficult to solve it with a particular mathematical equation.With the rapid development of deep learning technology,people begin to use convolutional neural networks to train and fit this many-to-one relationship,which significantly improves the quality of super-resolution images.At present,there are many methods to solve the problem of image super-resolution.But most of the network structures have a large amount of parameters and model calculation.Since the actual industrial application level is difficult to reach the computational power of experimental training,it will be difficult for this kind of method to be applied to real life.The research focus of this paper is how to maintain a good image super-resolution performance under the condition of relatively low model parameters and computational complexity.This paper proposes two research algorithms to solve this problem.(1)This paper proposes an image super-resolution algorithm based on feature separation and fusion network.Previous studies have shown that the performance of image super-resolution will be improved with the increase of network depth.However,it also increases the running time of the model greatly and makes the training difficult.The current situation that causes this problem is the lack of feature extraction ability of single-layer convolution,so researchers improve the model performance by adding the convolution structure.Based on this,it is particularly important to propose an effective network module.In this paper,a feature separation and fusion module is proposed,which learns the unextracted feature by making a negative residual between the feature extracted by single-layer convolution and the original shallow feature,and then adaptively fuses this feature with the feature extracted by single-layer convolution.This module is defined as the smallest structural unit.We use the idea of block,effective experimental analysis,and finally get the proposed feature separation fusion network.(2)In this paper,an image super resolution algorithm based on dense connection feature separation and fusion network is proposed.Based on the previous research method,the effectiveness of the feature separation and fusion module has been demonstrated.The research direction of this method is how to further improve the performance of the network without greatly increasing the model parameters.Because the flow of characteristic information between different levels can be achieved through dense connections,it is a very effective measure to learn and integrate the features of the previous levels in the process of network training.Then in the fusion module,we use a 1x1 convolution layer for feature fusion.It is also negligible in the consumption of parameters.Based on this,we propose a local cascade feature separation and fusion module and a global feature fusion module that constitute the core unit of our feature separation and fusion network.Finally,we have made a detailed comparison with some advanced algorithms in terms of visual effects and mathematical indicators.Our algorithm has the best performance.