Research On The Key Techniques Of Digital Face Skin Beautification And Overall Stylization

With the rapid spread of the social network and multimedia technology,the relationships between people are tightly connected.Face portrait,as each person's unique biometric identifier,is becoming more and more important.There is no doubt that everyone wants to have a beautiful and unique figure on social network,so the digital face beautification technology is attracting more and more attention.Digital face processing technology is a complicated problem involving multiple disciplines.Currently,there is no unified framework or evaluation criteria.This paper mainly designs and realizes digital face skin beautification and stylization.The main contributions of this paper as follows:(1)In order to avoid unreality by excessive operations,we propose an automatic facial flaw detection and retouching via discriminative structure tensor algorithm.First,a non-linear structure tensor with saliency model is exploited to discriminatively and automatically detect the significant facial flaws.Then,a Gaussian skin model constructed in YCbCr space with the OSTU operation is utilized to precisely mark the facial skin region where the mouth,eyebrows and nostril parts are excluded.Finally,we improve the exampler-based inpainting algorithm by tensor structure,which is used to retouch the detected flaws.The extensive experiments have shown its effectiveness,and the retouching performance is visually pleasing in comparison with state-of-the-art counterparts.(2)For portrait stylization,this paper utilizes a more popular deep generative model,Generative Adversarial Network,as the backbone for the framework.Specifically,we use the Generative Adversarial Network to learn the invariant representation of the face image,and take advantage of maximum mean discrepancy to minimize the feature discrepancy between face image domain and style image domain.The whole network is composed of generative network and discriminative network.The generative network,designed as encoder-decoder structure,aims to learning the fusion of face feature and style feature.The discriminative network takes charge of identifying face image,which resists the loss of identity information during the stylization process.When the training done,the proposed network can stylize the face image only by the generative network in real time.More importantly,the generated stylized face image would not be affected by the structure information of the provided style image.