The Research Of Transformation Of Face Pose Based On Improved Conditional Generative Adversarial Network

Benefiting from the rapid development of deep learning methods and the easy access to a large amount of annotated face images,unconstrained face recognition techniques have made significant advances in recent years.Although surpassing human performance has been achieved on several benchmark datasets,pose variations are still the bottleneck for many real-world application scenarios.In 2014,Ian Goodfellow opened up new ideas for image generation and face gesture transformation by proposing Genarative Adversarial Network(GAN).This paper adopts the method of conditional GAN,and improves the network from three aspects to solve the problem of face gesture transformation.Since conditional GAN will face the problem of gradient disappearance and network crash under the optimal discriminator,the Wasserstein distance is used instead of the JS divergence for calculation,and the network structure is improved as well.In the meantime,the problems of parameter search,super-parameter selection caused by multi-layer mutual coupling in network training weight update and over-fitting caused by noise data interference,which exist in deep learning,will not disappear in this network.Therefore,this paper adds batch normalization and dropout regularization methods to the network.CAS-PEAL-R1 dataset experiment results demonstrate the effectiveness of the improved method.The stability problem of network has been solved by WGAN,but weight clipping method adopted by it brings weight concentration and gradient explosion.Based on WGAN gradient Penalty(WGAN-GP),this paper uses the Versoria function to optimize the penalty parameter,so that the penalty parameter is closely related to the distance between the real distribution and the generated distribution,so that the penalty term plays a better optimization role,and the experiments prove the optimization method works.The experimental results show that the network has poor image quality and low rank-1 recognition rate.This paper combines the pixel-level L1 loss function and the structural similarity loss function to improve image quality from two aspects: reducing image blur and improving the rationality of face structure.The experimental results demonstrate the effectiveness of the algorithm combining the pixel-level L1 loss function and the structural similarity loss function.