Research On Image Super-resolution Algorithm Based On Deep Learning

Image has become a very important carrier for information display and dissemination in modern life.Therefore,the research of related algorithms has been widely valued by scholars at home and abroad,and many important branches have been derived.As an important branch of computer vision,image super-resolution technology has become an international hot topic in recent years due to its broad application prospects in many industries.The basic task of super-resolution technology is to reconstruct the corresponding high-resolution image or video from the original low-resolution image or video,which is a very challenging ill-conditioned problem.In recent years,deep learning technology represented by Convolutional Neural Networks(CNN)has made breakthroughs in many applications in the field of computer vision.The CNN-based super-resolution algorithm surpasses other classical algorithms in reconstruction effects because it can fully learn the implicit nonlinear relationship between low-resolution images and high-resolution images through training.In view of the shortcomings of several typical CNN-based image super-resolution algorithms,this paper proposes corresponding improvement strategies and verifies the advancement of the improved strategy through a large number of tests.The main innovations of the paper are as follows:Firstly,a method for estimating the parameters of degraded kernel based on grid search is proposed.Firstly,an imaging calibration system is designed and the imaging information in real scene is collected.Subsequently,the image blur degree quantitative analysis index is optimized,and based on this,the best image degradation is inferred by grid search.Finally,by training the "non-blind" super-resolution model ESPCN?NB under multiple kernel parameter settings,it is confirmed that the optimal kernel parameter setting does help to improve the reconstruction performance of the super-resolution model.Secondly,an image super-resolution algorithm with variable degradation information fusion is proposed.The algorithm is mainly designed for high resolution reconstruction of low resolution images under different degradation conditions.The super-resolution network with variable degradation information fusion is trained by taking the low resolution image under different degradation conditions and its corresponding optimal degraded kernel parameter as input.The test results show that the proposed super-resolution model can reconstruct the high-frequency details of low-quality images more accurately,especially for low-quality images with high degree of blur.The model has more advantages than other classical models.Thirdly,a multi-frame adaptive fusion video super-resolution algorithm is proposed,which mainly solves the problem of image registration in the traditional fixed frame number video super-resolution algorithm.This paper first proposes an improved video super-resolution algorithm for multiple sets of fixed frame fusion.Subsequently,a multi-frame adaptive fusion video super-resolution algorithm that dynamically configures multiple sets of fixed frame weights is proposed.The test results show that the proposed algorithm is better and less robust to low-quality video super-resolution reconstruction with large fluctuations in content information,and can effectively suppress flicker.