Research On Image Super-Resolution Algorithms Based On Deep Learning

Image Super Resolution(SR)is an important type of image processing technology in computer vision and image processing.It aims to enlarge the low-resolution image(Low-resolution Image,LR)by preserving rich details(such as colors and edges)and reduce interference(such as blur and artifacts).It can be used for different image processing tasks,including satellite imaging,military,medical imaging,and image magnification in mobile phones and tablets.Unlike the method of improving image quality by improving hardware conditions,the SR algorithm has the characteristics of low cost and wide applicability,so it has been paid attention by academia and industry.The SR method generally assumes that the LR image has been down-sampled from an unknown High-resolution Image(HR)through a fixed "ideal" down-sampling kernel(Bicubic down-sampling kernel).But compared with the LR data set synthesized in this way,this kind of situation rarely occurs in real LR images.When the assumed down-sampling fuzzy kernel deviates from the real kernel,the performance of the SR method will be greatly reduced.In view of the single problem of fuzzy kernel hypothesis in the process of degradation,this paper extracts the real fuzzy kernel in the mobile phone image set through Matlab code to degrade the HR image into a picture close to the real scene.At the same time,the fuzzy kernel is input into the network as a priori information of the image,and the use of Spatial Feature Transform(SFT)allows the network to better use the fuzzy kernel information to reconstruct the LR image.This paper tests the proposed model on the test set using Gaussian fuzzy kernel and real fuzzy kernel synthesis,and compares it with the current mainstream SR model based on deep learning.The experiment proves that the model has good evaluation indicators and visual effects.Aiming at the problems that the fuzzy kernel training set does not have diversity,the model is practical and poor in universality,etc.,this paper uses Generative Adversarial Networks(GAN)to optimize the model,the extract is carried out on the real fuzzy core based on the training of GAN.Then,the trained GAN model is used to generate samples that are close to the real fuzzy core,and the samples are used for model training to enhance the universality of the model.At the same time,the fuzzy kernel estimation module is added to the SR network model to replace the preprocessing step of extracting the fuzzy kernel to improve the practicality of the model.The obtained model is also compared experimentally on the test set.The experiment proves that the model can improve the universality of the network and has a better effect on quantitative indicators and visual effects.Aiming at the problem of complex model and slow convergence speed,the model is improved at last.Through the local jump connection,the fuzzy kernels in the shallow layer and the middle layer are merged with the image features in the reconstruction network,and the obtained image information is transferred to the deep network.In order to reduce network parameters while ensuring the effect,the number of network layers of the model is reduced.By controlling other variables through experiments,the same experiment as the third chapter is carried out,which proves that the improved method can accelerate the convergence of network and reduce the complexity of model under the premise of guaranteeing the reconstruction effect.