Face Super-resolution Via Densely Connected And Weighted Learning Network

Video Surveillance has made outstanding contributions in maintaining public security and cracking down on illegal crimes.Investigators can timely obtain the face image of suspects through Video Surveillance.But in the actual monitoring,the suspect is far from the camera and the resolution of the surveillance camera is so limited,which makes difficult to provide effective information.Face super-resolution(SR)is one of the technology aiming at solving the problem of reconstructing a high-quality and highresolution(HR)face image from the given low-quality and low-resolution(LR)face images.In recent years,many methods have been proposed and improved well progress,but even the latest deep learning methods have some drawbacks.On the one hand,these algorithms do not make full use of the face information,and image information is only transmitted in adjacent feature layers,thereby achieving relatively-low performance.On the other hand,the existing algorithm ignores the difference of the mapping functions in different regions.A single network is difficult to deal with all areas when rebuilding.The above two problems make it difficult to reconstruct a clear face,especially in the local area where there is a serious lack of detail.In order to improve the reconstruction effect of the face SR algorithm,we firstly address the problem that the existing algorithms do not make full use of the face information.In this paper,we propose a novel Multi-level dense residual network to address this problem in image SR.Then,in order to solve the problem that the mapping relationship between faces in different regions is different,inspired by adaptive algorithm and weighted learning algorithm,we propose to guide the reconstruction of face SR algorithm through weighted learning strategy.In particular,the weighted learning branch can explore mapping functions from LR to HR images in different regions,and distinguish different regions by the similarity of the mapping functions.The generator module can then be used to make a specific illusion on the selected area for better reconstruction results.Finally,we find that the pixel loss of the existing algorithm will lead to the relative blur of the reconstructed image,which difficult to satisfy the subjective feeling.To solve this problem,we add a discriminant network to constrain the data distribution between the reconstructed image and the real image,and improves the subjective effect of the reconstructed image.The experiment demonstrates the superiority of face SR algorithm based on densely connected and weighted learning network,compared with the state-of-the-art methods.The average PSNR of the reconstructed face images is increased by 0.61 dB,and the average SSIM is increased by 0.03.It's shown that our algorithm is of great significance for the application of face SR.