Super-Resolution Reconstruction Of Single Image Based On Convolutional Neural Network

In this era of continuous progress of information technology,people have higher and higher requirements for the resolution of images.As a technology to improve image resolution through software algorithms,super-resolution reconstruction has the advantages of low cost and good effect,and it plays an important role in many fields such as medical imaging,satellite remote sensing,and security monitoring.With the development of deep learning,convolutional neural networks are introduced into the field of super-resolution reconstruction.Because of its strong ability to learn and express complex data,it greatly improves the quality of image reconstruction.For this reason,this paper proposes two super-resolution reconstruction algorithms based on convolutional neural networks,the main research contents are as follows:(1)In order to solve the problem that correlation information between characteristic channels is ignored in many super-resolution reconstruction methods and information loss in network data transmission,a channel attention and residual concatenation network model for image super-resolution is proposed.The network is mainly composed of multiple residual concatenation groups connected in series.In each residual concatenation group,the input characteristics of the module will be concatenated and fused with the output of each subsequent layer to improve the information flow between different network layers and ensure that the characteristic information is not lost and fully obtained during the transmission of the network;At the same time,the attention mechanism is used to adaptively treat the feature channels fused in the module differently,which improves the reconstruction effect of the network on the high frequency information such as the details and textures of the image;In addition,the shallow features extracted from the front part of the network are transferred to each residual concatenation group through the fusion node for feature fusion,which makes full use of the feature information in the low resolution image.Experimental results on different benchmark data sets show that the model can improve the quality of image reconstruction.(2)In view of the different feature categories extracted from different receptive fields in convolutional neural networks,a multi-scale attention mechanism fusion super-resolution reconstruction network model is also proposed.The network uses two convolution kernels of different sizes for multi-scale feature extraction;At the same time,the improved attention mechanism is used for channel correction of the extracted features of two different levels to enhance the ability of the network to extract high-frequency information;Through multiple correction and fusion of feature information of different scales,the capability of feature extraction is improved;The network uses global residual learning to quickly transmit lowfrequency information and stabilize the training process.Experimental results show that the network has a good reconstruction effect.