Single Image Super-Resolution Based On Convolutional Neural Network

With the continuous development of society,it has entered an era of informatization.Among the large amount of information contacted every day,image information occupies a large part,so people continue to pursue high-quality images.The super-resolution reconstruction method,because it is a software-based method,has the characteristics of not involving hardware and strong universality,making it gradually become a research hotspot in the field of image processing,and in practical applications,image super-resolution reconstruction Technology is also widely used in medical,military and remote sensing fields.The main task of image super-resolution reconstruction is to obtain information from low-resolution images,and then reconstruct high-resolution images with higher resolution.The methods of image super-resolution reconstruction can be divided into two categories,which are reconstruction-based algorithm and sample learning reconstruction algorithm.For several years,the super-resolution reconstruction algorithm based on deep learning in sample-based learning,because it can process more data and can automatically extract features,establish a mapping relationship from low-resolution images to high-resolution images And compared with traditional methods,its reconstruction effect is far superior,so that the super-resolution reconstruction algorithm based on deep learning has gradually attracted people's attention and become a research hot.This article first introduces the development background,research status and common evaluation methods in the field of super-resolution reconstruction algorithms,and then introduces the basic knowledge of convolutional neural networks and some classic super-resolution in the field of deep learning.Rate reconstruction algorithm.Later,after analyzing these classic algorithms,it was found that the model depth of these algorithms is continuously deepening.At the same time,because the traditional convolution method is not efficient and contains a lot of redundant information,the complexity of the model is getting higher and higher.Aiming at these problems,this paper proposes a new residual structure,which adds an efficient convolution Octave Convolution.This convolution method divides the feature map into two parts,high frequency and low frequency,thereby compressing the information.At the same time,in order to better extract features,a channel attention mechanism is added to the residual structure.Subsequently,a model for super-resolution reconstruction of single-frame images based on the residual structure is proposed.The purpose of the model in this paper is to reduce the model's redundant information,reduce the amount of model calculation,and enhance the effect of image reconstruction.After a large number of public data set training,then use PNSR and SSIM evaluation methods in the field of super-resolution reconstruction commonly used verification set for verification.Experimental results prove that the algorithm in this paper can achieve good results while reducing the amount of model calculation.