Style-based Generative Adversarial Networks For Image Data Augmentation

In recent years,the rapid development of artificial intelligence technology has led to earth shaking changes in people’s clothing,food and transportation,especially the breakthrough of relevant technologies in the field of computer vision.From the most widely used face recognition technology to driverless,virtual reality technology and so on,we increasingly feel that research in the field of computer vision is always developing in a more convenient and diversified direction.While these excellent algorithm models can be applied in real life,which is inseparable from the support of computing power resources and data resources.Although major enterprises,universities,open-source communities and scientific research institutions have successively released open-source data sets for relevant tasks,with the increasing complex application scenarios,these open-source data sets are still far from massive data.Because collecting data requires a lot of time,money and personnel costs,this problem can be solved if data can be produced by technical means.At present,there are many ways of image augmentation in the industry.For example,adding Gaussian noise to the image,randomly clipping or masking the image,randomly rotating the image or enhancing the color of the image.These methods have been widely used in the training stage of various models.Nevertheless,the images generated by these methods are still lack of diversity and richness.From the perspective of the subject in the image,this transformation is actually no different from the original image.The proposal of style-based generator architecture for generative adversarial networks(Style GAN)gives us a new idea,that is to generate images through Style GAN.We combined the differential privacy algorithm with Style GAN.Our extensive experiments show that the image generated based on this method can well hide the identity information of face in the original image,so as to produce diverse and rich images.The main research contents of this paper are as follows:1.Because the Style GAN is the main method used in this paper,this paper studies the architecture of the model,training skills,the shortcomings,and evaluation metrics.At the same time,it briefly analyzes the downstream tasks based on the Style GAN,so as to pave the way for the following experiments.2.The method of combining differential privacy with Style GAN is adopted to solve the compromise between changing the facial features of human face and hiding the identity information in the Style GAN.Model training is carried out based on the large open-source face data sets,and the quality of the generated images is evaluated by using Frechet perception distance(FID)metric.Then the face identity clustering algorithm model is used to evaluate whether this method can well hide the identity information of the original training images.3.Due to the results of the style-mixing strategy proposed in Style GAN can only be obtained in the prediction of the model,the process is uncontrollable,and the style-mixing of the target images cannot be carried out one-to-one.So a strategy to fine-tune the intermediate latent codes of the target images is proposed and we reconstruct a one to achieve the controllable style-mixing of the target images.The image generated based on this method has no sense of conflict.4.This paper studies the series models of Style GAN,and we focus on the shortcomings of Style GAN,as well as the improvement and promotion of subsequent series models,which leads more researchers to think and look forward to relevant research.At the same time,since the self-attention mechanism is good at capturing long-distance dependence,and it can extract and activate important features through contextual information.Therefore,we introduce the self-attention mechanism into the generator of Style GAN to produce more realistic face images.