Research On Single Image Super-resolution Method Based On Hierarchical Network Structure

With the rapid development of remote sensing,surveillance,and imaging technology,people are paying more and more attention to obtain clear and high-quality images.However,it is expensive to obtain high resolution images directly through advanced hardware.Moreover,it is common to improve quality of low resolution image by algorithms,limited by the storage space and transmission bandwidth.Therefore,the image super-resolution has become a very valuable research project.Compared with sequence image super-resolution,single-image super-resolution has a wider range of applications.The focus of this paper is on single-image super-resolution technology.We find that the learning-based image super-resolution method is generally better than the interpolation-based one,after analyzing the current popular algorithms.However,the learningbased approach still does not yield satisfactory results for large up-sampling factors.On the other hand,for multi-scale super-resolution tasks,the strategy adopted by most methods is to learn multiple times to obtain mappings of low-resolution images to high-resolution images.This paper proposes and improves a new method of single image super-resolution.Firstly,a hierarchical progressive deep neural network structure for single image superresolution is proposed.The structure starts with the image pyramid idea and breaks down the complete task into multiple subtasks.Each subtask can be done separately by a single unit network.Multiple unit networks are cascaded to form an entire network.In this paper,structures with skip connection properties such as local residuals and dense connections are used to improve the efficiency of information flow transmission and avoid gradient disappearance.Since the objective indicators such as PSNR can not fully reflect the human visual perception,this paper proposes a corresponding discriminator based on the hierarchical progressive network.It constitutes a hierarchical progressive generative adversarial network for image super-resolution.Super-resolution images that are more visually sensible can be obtained through the network.Finally,test the network using data sets such as set5,set14,and BSD100.Compared with other super-resolution technologies,the PSNR of the reconstruction results of the hierarchical progressive network on each test set is 0.25 dB ~ 2.35 dB higher,and the SSIM is 0.003 ~ 0.053 higher.The generative adversarial network can restore the details of the image better for large up-sampling factors.