Research On Super-resolution Image Reconstruction Based On Deep Learning

Image is one of the important carriers of information transmission.In real life,about 70% of the information is obtained through vision.The resolution of the image reflects the richness of the information contained in the image.So how to obtain high-resolution images efficiently is one of the difficult problems to be solved.Increasing image resolution from a hardware perspective will face difficulties in research and high cost.Image super-resolution reconstruction is a kind of method to improve image resolution by software.It has high cost performance and has wide application value in many fields.Therefore,it has become a hot spot for experts and scholars.This paper starts with the image super-resolution algorithm researched by previous scholars,expounds the research status of the technology,and reproduces several image super-resolution network models based on deep learning,analyzes and integrates the advantages of existing network models,and combines the basic theoretical knowledge deep learning research for single-image super-resolution techniques.The main work of this paper is as follows:(1)Find information on image super-resolution technology and explain the current state of image super-resolution technology.Reproduce image super-resolution technology multiple implementation algorithms,and focusing on super-resolution algorithms based on deep learning.Analyze the advantages and disadvantages of each network model and compare the experimental results.(2)For the low utilization rate of the extracted feature information in the current image super-resolution reconstruction network,the texture details of the reconstructed super-resolution image are slightly insufficient,and the texture part recovery is difficult.In this paper,a high-frequency information of low-resolution image is constructed as a compensation information model for reconstructed image,and formed feature compensation Deep Convolutional Neural Network(DCNN)image super-resolution reconstruction algorithm.The original low-resolution image is interpolated and transmitted into the U-shaped network to extract the texture information of the image,and then added with the image reconstructed by the DCNN model to form texture feature compensation,and the final high-resolution image is obtained.After experimental comparison,the super-resolution image reconstructed by this algorithm can obtain higher objective data of image quality evaluation and stronger edge feature.(3)For the current image super-resolution algorithm,the single-chain network structure is used to extract the feature information of the low-resolution image using a single-scale convolution kernel,which easily leads to the omission of detailed information.In addition,in order to obtain better image super-resolution reconstruction effects,and the accompanying gradient disappearance problem will make the training time prolonged and the difficulty is increased with the network model constantly deepned.This paper proposed a multi-scale intensive residual network model based on GoogleNet,residual network and indensive convolutional network.In this model,three different scale convolution kernels are used to convolve the input low-resolution image,and the underlying features under different convolution kernels are collected,so that the detailed information on the low-resolution image can be extracted more,which is beneficial to recover image.Finally,the feature information extracted by the three convolution kernels after fusion is reconstructed to obtain a clear image.Experimental data and renderings prove that the proposed algorithm can recover the edge and texture information of low-resolution images better than the current mainstream super-resolution reconstruction algorithm.