Research On Blurred Image Restoration Based On Generative Adversarial Networks

With the accelerated breakthrough and widespread use of Internet,big data,supercomputing,Internet of Things,and communications,artificial intelligence has entered a new phase.Deep learning as the core technology of artificial intelligence plays an important role.And GANs has become a dark horse and emerged as an irresistible trending in the deep learning community since its inception in 2014.At present,GANs have made great achievements in the field of image processing.GANs,as an excellent generative model,has been widely used in image generation,patching,denoising,style conversion,and natural language processing.Image restoration has always been a research hotspot and difficulty in the field of image processing.Image degradation problems such as the motion blur,Gaussian blur,salt and pepper noise,and mosaic blur mentioned in this paper seriously affects people's visual experience and scientific research and industry applications.Therefore,research on deblurring blurred images is of great research value and significance.For this type of blurred image,this paper proposes a research model based on the generative adversarial networks of image restoration.The main innovations of this article are as follows:(1)This paper uses the current hot Generative Adversarial Networks and establishes a data-driven end-to-end image deblurring model.The model is semi-supervised,requiring pairing data between the blurred image and the original image during the model training stage.After the model training is completed,the paired image data is no longer needed,and the blurred image is input to generate deblurred image as the output.(2)Due to the neural network and other methods,the loss of image is calculated based on pixel value.such as L1,L2 loss,which is not related to human perception of image quality.Hence,this paper proposes a content loss function based on similarity of image structure SSIM.SSIM can capture the complex features of human visual system(H VS).(3)Different from the traditional deblurring method that can only deal with single blurred image.The method proposed in this paper can handle a variety of blurred images under the condition of sufficient model training.For the unknown kind of blurred image as input,this method can be directly used to deblurre,without the need to prejudge the fuzzy type and select the corresponding image restoration method,to achieve a complete "end-to-end" intelligent image deblurring process.