Research And Application Of Image Super-resolution Reconstruction Based On Deep Learning

Image is one of the important ways for people to perceive information.The higher the resolution,the richer the detailed information.With the continuous development of science and technology,people's demand for high-resolution images is increasing.However,due to many factors,the quality of images captured by various types of imaging equipment is often low.Therefore,in order to better solve this problem,it has produced super-resolution reconstruction technology.This technology is mainly a method of reconstructing the corresponding high-resolution image from the original low-resolution image using a related algorithm.In recent years,the improvement of deep learning technology has made the super-resolution reconstruction technology achieve breakthrough results that cannot be compared with traditional algorithms.Therefore,based on the current mainstream network models,this paper deeply researches some of the problems in the reconstruction process and gives the corresponding scheme to further improve the quality level of the reconstructed image.The main research contents of this paper are as follows:(1)Firstly introduces the research purpose and significance of super-resolution reconstruction and the current status of research at home and abroad,and then outlines some of the theoretical knowledge involved in this subject,and classifies and describes the related classic algorithms of super-resolution reconstruction,which focuses on learning-based reconstruction,and briefly outlines its advantages and disadvantages.In order to facilitate the experimental comparison of different algorithms,this paper also briefly summarizes the image quality evaluation system.(2)Aiming at the problems of current deep learning-based image super-resolution algorithms,such as the feature extraction scale is single and the reconstructed image texture is blurred,this paper proposes a super-scale reconstruction network model based on multi-scale recursive network.The model constructs a recursive learning network by cascading multiple multi-scale feature mapping units.Each mapping unit is connected by a set of feature extraction layers of three scales,an information fusion layer,and a feature mapping layer.The network directly takes the original low-resolution image as data input,and uses recursive learning to deeply extract the detailed features of the low-resolution image to obtain the corresponding feature map.Finally,the sub-pixel convolutional upsampling method is used to reconstruct the high-resolution image.The experimental results show that the network has achieved better super-resolution effect,subjectively the image texture is clear and the edges are sharp,and the objective data indicators on the standard test set are also higher than the existing mainstream algorithms.(3)In order to improve the resolution of medical images,this paper builds a super-resolution reconstruction network model applied to CT images to solve the problem of the existing model's structure and unity of multiples.Drawing on the idea of residual network,we construct multi-scale residual blocks by combining global residual learning with multi-scale information features,and use multiple residual blocks to sequentially learn the residuals feature between low-resolution CT images and high-resolution CT images,and finally the residual feature map is fused with the input CT image to reconstruct the required high-resolution image.The final experimental results show that the subjective rendering of the network reconstruction can more clearly reconstruct the detailed features of human organs,and higher objective data indicators further demonstrate the effectiveness of this model,and only need a set of parameters to simultaneously Supporting multiple image magnifications improves the practicability of the algorithm to a certain extent.