Research On Face Super-Resolution Reconstruction Technology Using Generative Adversarial Network

Face Super-Resolution(Face SR)refer to a technology which employing Low-Resolution(LR)face images and facial inherent attribute to reconstruction High-Resolution(HR)face images.As an important part of face image processing technology and image super-resolution(SR)technology,Face SR has always been the focus in the field of image vision.In the actual scene,due to the influence of lighting,facial expressions,shooting environment and other factors,the HR information loss of face images is serious,and the reconstruction of facial features has always been a difficult problem in Face SR research.In recent years,the deep learning-based Face SR has attracted wide attention of scholars because it can effectively use a priori information,and has made important progress in improving the reconstruction quality of face images;The emergence of the new Generative Adversarial Network(GAN)also provides an effective technical means for Face SR to enable the related Face SR method to reconstruct natural,high-quality face images in the absence of serious information.In this context,based on the GAN network,this paper studies the Face SR technology based on the deep learning method.The specific work is as follows:(1)Aiming at the problem of effective expression of face prior information,this paper proposes a method based on deep network for the expression of face component semantic prior.The method employs the Semantic Segmentation Probability(SSP)technique to define the probabilistic attributes of the face image region components and the expression of the prior information in Face SR,and then utilizes the defined face component attributes and the component texture and edge information contained in the semantic prior method,the texture and edge of the face image are reconstructed in a targeted manner to restore the fine-grained,high-quality HD face images.The experiments have shown that the expression method surpasses other prior expression methods in the existing mainstream evaluation methods and visual effects of images.(2)Based on the semantic prior,a high performance Face SR method based on GAN(CSP-GAN)is proposed in this paper.The method designs a new GAN network architecture for Face SR,which enables the generation network and the discriminant network to make full use of the previously defined face semantic prior information.In the generation network,a Component Semantic Prior(CSP)is designed,which can effectively combines the face prior information including texture details and shapes by using the affine transformation technique.The component semantic information and feed-forward modulation method of image features improve the training method of generating networks and improve reconstruction performance;In the discriminant network,a semantie discriminating module is designed to determine the semantic categories and properties of the images,and then through component semantic prior feedback to generate high quality face eomponent shapes and fine-grained eomponent textures.Experiments show that the proposed CSP-GAN can effectively reconstruct high-quality face images by utilizing face information in a priori information.In summary,the GAN-based Face SR proposed in this paper can effectively utilize the semantic prior information extracted from the training data to achieve high-definition face image reconstruction,which has a good application prospect.