Research On Super Resolution Reconstruction Algorithm Based On Deep Learning

With the development of technology,image information,as one of the most important information,has irreplaceable significance in production and life.The resolution of the image is usually used to reflect the clarity of the image,but high resolution images cannot be obtained in some conditions,and super resolution techniques can be used to reconstruct high resolution images.The aim of single image super resolution is to reconstruct the high resolution image from its corresponding single low resolution image,super resolution has been widely used in satellite imaging,medical image processing,security monitoring,image enhancement and image compression.Super resolution reconstruction based on deep learning is a hot topic in current,and it also achieves a good reconstruction result.However,the better reconstruction result of super resolution usually requires a large network model,it also brings some problems such as the complexity of the algorithm and the storage space of the storage of network model are also increasing dramatically.Some devices such as embedded devices and mobile devices that are limited in terms of computing,volume,power consumption,it is not possible to efficiently apply a large network model in these devices.This paper discusses how to achieve good super resolution results with smaller convolutional neural networks to facilitate the deployment the super resolution model in low-power,low-computing-capacity embedded devices or mobile devices expediently.The main research contents of this paper are as follows:(1)This paper applies the knowledge distillation in the super resolution problem innovatively,a student network for super resolution with shallow network and few parameters to learns the knowledge of a teacher network for super resolution with deep network and large parameters and improve the reconstruction results of student network for super resolution without changing its parameter quantity and network structure.This paper proposes a super resolution student network model based on Mobile Nets,a lightweight network,so that it can effectively run in resource-limited devices.Using Peak Signal-to-Noise Ratio(PSNR)and Structural SIMilarity index(SSIM)as the evaluation indicators.And compare a variety of different methods for extracting knowledge,and select the best way.The student network using the knowledge distillation method has a peak signal-to-noise ratio increase of four public test sets at upscaling is 3 compared to the student network without the knowledge distillation method were 0.53 d B,0.37 d B,0.24 d B,and 0.45 d B,respectively,and its time and memory consumption will not increase.(2)The student network cannot toughly learn the knowledge in teacher network,thus,the difference of reconstruction between student network and teacher network is still exists.The former method focuses on peak signal-to-noise ratio and structural similarity index as the evaluation indicators,so optimizes the pixel-level loss function which leads to less texture detail information of reconstruction high resolution image.To deal with this problem,based on the above research,the trained student network is trained by generative adversarial network to further enhance the super resolution reconstruction results of the lightweight network,using this way to obtain high resolution image that can accord with the subjective feelings of the human eye.The experimental result shows that this method can effectively improve the visual effect of super resolution reconstruction.