Research On Single Image Super-Resolution Based On Deep Learning

In recent years,with the development of computer vision processing technology,computer vision processing has been applied to various fields.Therefore,people have higher and higher requirements on image quality.However,in some application scenarios,it is often difficult to obtain high-quality images due to the influence of imaging equipment,imaging environment and other factors.This will greatly affect the effect of image processing.In order to solve this problem,the image super-resolution reconstruction technology based on algorithm comes into being.The image super-resolution reconstruction technique is to reconstruct a low-resolution image into a corresponding high-resolution image through an algorithm.The technology is widely used in medical images,satellite remote sensing images,and security monitoring,with important practical application value.The traditional super-resolution reconstruction methods are interpolation and prior information based methods.Although this method is relatively simple to implement,the refactoring effect is not good.Most of the existing image super-resolution reconstruction methods are based on deep learning technology,but there are still some problems such as too smooth reconstruction effect,difficult network model reproduction,unstable training and so on.This paper focuses on the improvement of deep learning network model for image super-resolution reconstruction.The following work was mainly carried out:First of all,this paper sorted out the development of deep learning based reconstruction method,systematically analyzed and summarized the differences and connections between different methods,and analyzed and compared the reconstruction effects of several major deep learning networks in the main data sets.Based on the systematic comparison and analysis of the deep learning network model used for image super-resolution reconstruction in recent years,this paper proposes an image super-resolution reconstruction method based on deep residual dense network(DRDN).Specifically,it includes:1)taking dense connection network as the basic unit,fully integrating and utilizing the hierarchical feature information extracted from the convolutional layer of each layer in the network to provide richer feature information for the final reconstruction;2)aiming at the problem that deep network is difficult to train convergence,and inspired by the idea of residual network,the identical connection is introduced between the output layer of each dense network unit and the input layer of network for residual learning.3)the residual learning results of each dense network unit in the network are taken as the intermediate reconstruction results of the network.Finally,the weighted sum of intermediate results is achieved by introducing the reconstruction layer with convolution kernel size of 1 1 to obtain the final reconstructed image.Experiments show that the deep residual dense network proposed in this paper achieves good reconstruction effect,and at the same time ensures fast convergence speed and good training stability of the network model.Then,on the basis of deep residual dense network model,the method of acquiring training data is carried out.When most existing models acquire training images,they usually carry out bicubic interpolation through high-resolution images to obtain corresponding low-resolution.This limits the reconfiguration performance and generalization ability of the refactored network to a certain extent.In this paper,a common noise model is introduced and combined with the bicubic interpolation,the high-resolution image is processed comprehensively.A more complex degradation model is established to simulate the low-quality images in real application scenes more realistically.The DRDN model proposed in this paper can achieve good reconstruction effect.Combined with the improvement of training data degradation method,the reconstructed network has good generalization ability.