Image Super-Resolution Based On Convolutional Neural Network

Super-Resolution(SR)is a process of recovering corresponding high-resolution images by means of software algorithms using one or more low-resolution images.It is widely used because of its flexible operation and low cost.Convolutional neural networks have achieved good results in SR reconstruction compared to traditional methods.However,the network structure of these models is very shallow,and the perception fields of convolution kernels are not large enough,which leads to the failure to make full use of the context information of the image to infer the missing high-frequency information during the reconstruction process,and the complex texture detail reconstruction effect is not it is good.For this reason,these models failed to reconstruct complex texture details.In addition,blindly increasing the depth of the convolutional neural network will increase the network parameters,resulting in extremely high computational complexity,increasing the training difficulty of the network,and failing to improve the network effectively.In addition,as the depth of the network increases,the image features gradually disappear during the propagation process,and the low-resolution image features are not fully utilized.The main research contents and innovations of this paper are as follows:(1)In order to solve the problem that the perception field of shallow network convolution kernel is not big enough to make the most of image context information to reconstruct high quality image,a deep residual network for image super-resolution reconstruction is proposed.Learning global residuals by the skip connection while learning local residuals in each residual block.Based on this,while the network structure is deepened,the training speed of the network is also improved,and the perception field of the convolution kernel is also fully expanded.The network makes good use of image context information.(2)In order to solve the problem that image features disappeared during transmission and are not fully utilized,this paper proposes a multi-scale residual block composed of multi-scale residual units,which can adaptively detect image features of different scales.In this paper,the multi-path connection method is used to replace the chain connection of the common network.The original image features are transmitted to the deep layer of the network so features extracted from the input image are more fully utilized to further solve the problem of feature disappearance during the propagation process.Meanwhile,the multipath connection helps the back propagation of the gradient during training.Recursive learning increases network depth while controlling network parameters.In this paper,two new super-resolution reconstruction models based on convolutional neural networks are proposed.The experiments prove the improvement of network results by deepening the depth of the network,making full use of image context information,multi-scale extraction features and recursive learning.