Research On Intricate Face Recognition Based On Virtual Sample Generation

Face recognition was put forward and developed in 1950 s and 1960 s.It has become the most prominent and widely used identification technology in biometrics.Face recognition technology is a cross research in the field of computer vision and image processing.It has a broad application prospect and meets the practical application requirements.As a kind of biometric recognition technology,face recognition has the advantages of convenience,non-contact and non awareness to the masses compared with other recognition technologies in application occasions.In recent years,it has attracted the attention and research of University researchers and relevant enterprises.In today's technology driven,under the condition of controllable environment and user cooperation,face recognition system can achieve better recognition rate.However,with the widespread use of face recognition technology in practical applications,face images are often vulnerable to face pose changes and occlusion,which brings severe challenges to traditional face recognition research.One of the difficulties is that complex faces such as face posture changes and occlusion will make the facial features lack of feature information,which cannot meet the needs of recognition;the other is that there is no special database for complex face recognition due to the large posture changes and the lack of number of occluded face images in the current database.In order to solve the problem of large pose change,occlusion and other training samples in real images,this paper presents the advantages of constructing virtual samples and combining deep learning to extract deep feature information,to perform complex face recognition experiments.Based on the deep learning based on massive data,there are two main aspects of the current face recognition technology,which are the small number of occlusion picture data samples and the changes in pose and occlusion.This paper is divided into two parts: the first part is to use the virtual sample generation of style migration star network to solve the problem of insufficient image data such as large pose change and occlusion;the second part is to use the expanded data set of the generated virtual sample,and then combine the deep learning to realize the face recognition in the complex situation.The main research work of this paper is as follows:(1)starting from the two aspects of large posture change and occlusion,and comparing and summarizing the advantages and disadvantages of the current solutions,by generating a large number of virtual sample images with large pose changes and occlusion,it can reflect the change of the missing face image samples in the original database,the combination of virtual samples and original database can meet the requirement of deep learning for training image data of face recognition,and solve the problem of complex face recognition.(2)Aiming at the complex recognition experiment,it is necessary to construct and generate face images under complex conditions such as large pose change and occlusion similar to the training data set.a face attribute transformation based on style migration StarGAN network is proposed.By adjusting the structure of the model,the generated virtual sample image can achieve better results in both image quality and multiplicity without changing the basic character information.The structure of the model generates a virtual sample image with large attitude change and occlusion,and adds training samples to make up for the lack of image quantity.(3)In the detection and recognition based on Faster R-CNN,the Faster R-CNN is combined with the style migration based StarGAN network,finally,the virtual sample generated by the StarGAN network is combined with some image data of the original data set to construct a new large pose and occluded data-set.Using Faster R-CNN based on deep learning,it can achieve a higher detection rate and recognition rate,and it is more robust for face detection and recognition in different poses and occlusions in the face of sunglasses.Since the introduction of deep learning,facial recognition problems such as expression and lighting have been solved.The recognition rate has dropped rapidly due to the lack of a large amount of facial information caused by posture changes and occlusion.Therefore,exploring pose recognition,face occlusion and other fast and efficient identification problems in complex facial environments is of great significance for revealing large pose changes and occlusion recognition laws,and lays a solid foundation for the research and promotion of face recognition technology.By combining the virtual face images of sunglasses and posture changes generated by face attribute conversion based on the style transfer StarGAN network and Faster R-CNN under deep learning,complex face recognition of posture changes and sunglasses in real scenes is realized.However,there is no relevant theoretical basis for the face with attitude change,and the face with different angles selected in the experiment is biased,so the effect of generating the virtual face with attitude change on the contour is not good,which is the inadequacy of this paper,and is also the subsequent research direction.