Research Of Adaptive Facial Image Beautification Based On Generative Adversarial Networks

Face image beautification is an important research direction in the field of deep learning and machine vision,which is applicable to modeling design,face-lifting,digital image beautification and virtual character design,etc.Face image beautification is attracting more and more attention of researchers and has a wide application prospect,but there is still no unified face beautification method or evaluation criterion.Besides,most of existing beautification algorithms have their own advantages and disadvantages.Therefore,more people have turned their attention to the field of deep learning.In this field of study,Generative Adversarial Networks is one of the most promising and potential generative model,it has been mostly applied to computer vision tasks to generate samples of natural images,and can also be used in the field of natural language processing,music composition and video composition.This paper mainly designs and implements an adaptive face beautification model based on Generative Adversarial Networks,which is able to use real face images and given conditions to generate beautified face images.The main contribution of this paper are as follow: First,we study the application of Generative Adversarial Networks and its related work in the aspects of grayscale image colorization,image synthesis and terrain modeling.Then,we compare the results of different improvement methods and analyze the effect of parameters in these models.So we gain a deeper understanding and more practical experience of Generative Adversarial Networks.Second,we train a discriminative model to evaluate the beauty level of training samples,and then use the results of evaluation as the label of unlabeled samples.This model can help us to get enough labeled samples for training.Third,we design the architecture of face beautification model,and use the labels of training samples as the condition of generation to control beauty level of generated samples.To meet the need of human face beautification,we add some methods such as variational auto-encoder,random cropping,gradient penalty,etc.,to improve the beautification ability of the model and the stability of training.Forth,we analyze and evaluate ability of our face beautification model based on the experimental results to adjust the hyperparameters and network structure,so as to improve the quality of the generated samples.Then,we use other beautification methods and deep learning methods to postprocess the generated samples to further improve the beautification effect.The innovation points of this paper are as follow: first,we separate the convolution to decrease the relativity of inner parameters.Second,we apply gradient penalty and perceptual path length to stabilize the training process.Third,we use conditional restraints to control the feature of generated samples.Forth,we combine the pretrained VGG16,which can be seen as an encoder,and the generator of generative adversarial nets together to make a Variational Auto-encoder.Then,we connect the corresponding layers of generator and encoder by fromRGB and toRGB layer.The advantages of this model are as follow: the training process is stable and efficient,the operating procedure of user is simple and fast,the beautified samples are diverse but not monotonous,and the beauty level of these samples can be directly adjusted and controlled.