Face Super Resolution Based On Deep Convolutional Neural Network

Face super-resolution is a method to improve the resolution of face images.Compared to improving hardware equipment,it has the advantages of low cost and easy upgrading.This technology can be applied to video surveillance,face recognition,multimedia communications and so on.The deep learning algorithm has achieved great success in computer vision.General superresolution based on convolutional neural networks(CNN)achieved great results too.However,CNN is an end-to-end network,it's difficult to use facial priors in the network.Most of face super-resolution methods based on CNN use the networks similar to the general superresolution and they do not use the facial priors.At present,the good results on face superresolution based on CNN significantly benefit from the robust feature from the CNN.Based on the robust features from CNN,we present a two-stage algorithm for face superresolution,which take facial priors into account.Generating a face from a single CNN is not good enough,because a face contains many different details such as the big difference between eyes and mouth.The two-stage method proposed in this paper include two steps,outline restoration and detail restoration.First,we use the outline restoration CNN to generate the whole image.Second,we use the detail restoration CNNs to generate the patches which have important details.At last,we merge the images from the two step into one.We compare our method to some state-of-art methods.Experiences show our methods have better results and wider usages.In addition,we discuss the idea of designing the architecture of the detail restoration CNN.And we discuss the effect of the depth of the network from the feature map's perspective.At the end of this paper,we point out using detail restoration patches replace the outline restoration will lead to some trouble.And we proposed two method to get better results,one is using image smoothing,the other is using image fusion based on wavelet transform.