Research On Single Image Super-resolution Reconstruction Algorithm Based On Generative Adversarial Networks

The main approach which mankinds obtain information relies on vision,therefore image plays a crucial role in the process of information transfer.Compared with ordinary definition images,high-resolution images represent the recognizability of their detailed features has been greatly improved,which is beneficial to people to fully extract the key information from them and acquire corresponding knowledge.In the past few decades,super-resolution(SR)technology has been an extremely important research subject in the fields of computer vision and image processing and so on and has been widely used.In recent years,breakthroughs have been made in researches on various fields with the rise of deep learning technology.As a landmark generative model in deep learning domain,generative adversarial network(GAN)provides a reliable theoretical basis and algorithm support for the development of deep learning technology in computer applications and researches.This paper takes the image super-resolution technology combined with deep learning as the research background and makes following researches on single image super-resolution reconstruction algorithm which is based on generative adversarial networks:1.Investigate the typical image super-resolution reconstruction algorithms and classify them into two categories: non-deep learning approaches and deep learning approaches.And then make a comprehensive comparison of the characteristics of various methods,highlighting the superiority of deep learning approaches in image super-resolution reconstruction problems.Through the experimental results on standard datasets(Set 5,Set 14,BSD 100 and Urban 100)to confirm the advantages of deep learning approaches compared with the traditional methods.And meanwhile reflect the current problems and defects of super-resolution methods which are based on convolutional neural network(SRCNN)and very deep network(VDSR).2.Make a deep analysis of the advantages & disadvantages and application scope of generative adversarial network(GAN)for super-resolution reconstruction tasks.The experimental results on standard datasets(Set 5,Set 14,BSD 100 and MNIST)illustrate the super-resolution based on generative adversarial network(SRGAN)model has problems in the detailed features of generated images and image quality assessment(IQA)results.And make a discussion about the improvement directions and measures of above problems from three perspectives of loss function,network structure and evaluation index.3.Propose a class-information generative adversarial network superresolution(Class-info SRGAN)model.The advantage is that by introducing the feature class-information into the perceptual loss function and improving the optimization goal of the original super-resolution model,which makes the feature representation of generated results become more prominent.The experimental results on standard dataset(Celeb A,Fashion-mnist and Cifar-10)verify the superiority and effectiveness of the proposed method in super-resolution reconstruction tasks.4.Put forward to use the perceptual-distortion tradeoff theory in image restoration algorithm to explain the inconsistent performance between the subjective and objective evaluation results of existing SRGAN model generated images.And use the perceptual index(PI)to test the perceptual quality of generated results,thereby achieving a more objective and accurate evaluation of generated images.The feasibility and applicability of the hypothesis was verified by experimental results on standard datasets(Set 5,Set 14 and BSD 100).This paper enriches the research and application ideas of generative adversarial network(GAN)structure in the field of image super-resolution reconstruction which based on the work of SRGAN model and provides a reference in algorithms and models to design super-resolution models which can generate images possess a more ideal quality based on this model architecture.