Super-Resolution Reconstruction Of Face Image Based On Deep Learning

The role of video surveillance in the field of public security is becoming more and more important.However,one of the main challenges is that due to the limitations of the imaging device itself and the complex environmental impact of the public place,the resolution of the face image in the video is often low,so that the definition is poor and difficult to recognize.Therefore,how to reconstruct high-quality face images using image super-resolution reconstruction technology has become a research hotspot.The goal of this paper is to reconstruct a low-resolution(16*16 pixel)single-frame face image(or face video)into a high-resolution(64*64 pixel)single-frame face image(or face video).The specific research content is as follows:(1)Firstly,a variety of typical image super-resolution reconstruction methods are reviewed.Compared with traditional methods,deep learning based methods have greater advantages.Then,a variety of typical convolutional neural network models are analyzed.The deep extraction features are higher in order and the rebuilding ability is stronger.(2)For the single-frame face image,two image super-resolution reconstruction methods are proposed,which are based on the cascading residual network and the method based on generative adversarial network.The method of cascading residual network is used to accumulate the image features of each layer in the network to improve the reconstruction quality.Based on the method of generative adversarial network,the generator model is built based on the dense network,because the dense network uses parallel connection to effectively preserve the image details of each layer in the network.(3)For face video,a super-resolution reconstruction method of face video based on single-frame reconstruction network is proposed,which can reconstruct rich details of face image and obtain better reconstruction effect.(4)The influence of the loss function on the reconstruction effect is analyzed.The method of using the mean square error as the loss function can improve the peak signalto-noise ratio and structural similarity of the reconstructed image.Introducing the perceptual loss can enrich the details of the reconstructed image and improve the visual effect of the reconstructed image.