A Multi-View Face Generation System Based On Generative Adversarial Networks

The combination of information and industry is increasingly close,and the booming development of the medical beauty industry is not behind.The detection of head posture and the prediction of face Angle are conducive to face recognition and face reconstruction.Medical cosmetic surgery is an important part of preoperative analysis,most of good preoperative analysis can shorten the operation time,face precision analysis is relatively important,in view of the small clinic equipment is insufficient,or when the customer is busy or abroad,unable to participate in preoperative information gathering,is bound to affect the correct operation process,customer involvement will also reduce the surgery,the whole process of time and growth,is a set of complete analysis of generating system will bring considerable convenience.The problems of face correction,recognition and reconstruction are often interrelated,and face recognition can be seen everywhere in applications such as emotion detection,access control system and face payment.Good face correction can lead to better face recognition,good face reconstruction can keep good face information under effective recognition,and make face information prediction with reconstruction results.In this paper,an improved generation countermeasure network model is proposed to generate a two-dimensional face based on the two-way training generation countermeasure network,and a face generation and correction system which can achieve a three-dimensional face alignment correction is combined.The results are corrected by human eyes to generate a positive standard angle face.In this paper,a high-precision multi-view face correction and standard face generation system is developed by using the generation countermeasure network.By training face image files with different side angles and pitch angles,a single face image can be input to generate a face image with a specific standard angle.This paper comprises four major features.First,we recognize and extract faces of particular angles from input image sequences to construct two-dimensional(yaw and pitch)face dataset.Second,we optimize the architecture of GANs by the extended two-dimensional face dataset and shared weights and features.Third,we propose a strengthened training mode for single pose images to effectively retain particular pose information.Last,utilizing facial angle alignment,we could generate any angle faces according to the demands.We compare the generated images of standard angles with the ground truth and analyze the deviation of facial landmarks by the face recognition model of Facenet.Compared to the average L2 distance(0.820)incurred by the CR-GAN method,the proposed method achieves a much less L2 distance of 0.654.