Super-resolution Of Single Image Based On Deep Learning

Image super-resolution reconstruction is an important research direction in computer vision and can be applied in many fields in real life.In our work or daily life,due to equipment cost,technical limitations,the limitation of the grid transmit capacity and other factors,it is common that we can not get enough clear images.If we just rely on the improvement of the hardware equipment,it will not only cost a lot,but also be difficult to implement in some cases,such as the interference of human tissues in the medical imaging of human body.Therefore,the super-resolution technology of image is very important in some time.The technology of image super-resolution reconstruction has been applied in many fields,such as medical imaging,remote sensing imaging and surveillance video.The detail information of super-resolution image based on traditional methods is not perfect,the visual effect is also difficult to satisfy humans.In recent years,super-resolution technology based on deep learning has made many great achievements.However,there are still some problems,such as the noise problem in the reconstructed texture.Based on this,this article will work on the super-resolution technology which is based on Generative Adversarial Networks.On this major model,the main work of this paper is as follows:(1)We analysed the Convolutional Neural Networks and Generative Adversarial Networks including their models,training methods and advantages or problems in image processing.Such as when we extract image features,all the feature mapping are unified,image information is mixed,this may be the cause why reconstruction image has certain deviation or noise in some areas.In order to solve this problem,some novel methods are proposed,such as using Octave Convolution to replace the original Convolution layer,in this way,we can separate the high and low frequency information in the image,optimizie the processing method for detail and texture.(2)Also,we improved feature extraction network and perception loss.A feature extraction network for the current data set was added to the original single pre-trained VGG19 model to make the original feature extraction more targeted and improve generalization performance.(3)In addition,we studied the new activation function and optimization algorithm.The original super-resolution models basically used traditional activation function and optimization algorithm,such as Re LU family and Adam series.In recent years,researchers have been carrying on the unremitting researches.On this basis,this article used several new activation function and optimization algorithms proposed by the deep learning areas,such as Swish and Mish activation functions,some new optimization algorithms like Ranger and Novograd.This research conducted control experiments to compared their effects.(4)Finally,the practical and feasible improvement ideas wre applied to the image super-resolution reconstruction experiment and conducted the superresolution reconstruction experiment with the improved network to verify the effectiveness of the improved model by comparing the effect of the original model.According to the previous experience,experimental results were compared on the general test data sets Set5 and Set14.We used the objective indexes,PSNR and SSIM,to evaluate the results and showed the visual effects of image reconstruction.The experiment conclusions can verify that this improved super-resolution model can well deal with the problem of image superresolution reconstruction.