Face Recognition Based On Face Pose Reconstruction Technology

Face recognition is a multi-disciplinary biorecognition technology,which is a research hotspot in the field of pattern recognition and computer vision.The current face recognition technology has achieved good experimental results on the front view,but in actual life,the pose of the collected face image usually changes due to the environmental impact.The change of face pose is an important reason for the decline of face recognition rate,and is a technical bottleneck in the field of face recognition.The core idea of this paper is to eliminate the influence of attitude change on face recognition through face attitude reconstruction technology.However,due to self-occlusion,it lacks most of the facial features,so it is very difficult to reconstruct the facial pose.This paper focuses on the problem of face pose reconstruction and summarizes the current mainstream research methods.It divides face pose reconstruction into two aspects: face positive and face pose generation,and researches and improves this.The main contents of this study are as follows:(1)Establish a face database and face preprocessing.Use multi-task cascaded convolutional neural network for face detection and face alignment,affine transformation according to the coordinate position of face feature points,straighten the inclined face,to reduce the impact of complex background in the image on the face reconstruction experiment.(2)Face frontalization.Inspired by the research work of generative adversarial networks on the transformation of face color,hair and other attributes,this paper uses the face deflection angle as a pose attribute of the face to perform interchange training based on the generative confrontation mechanism,and proposes a condition-based Face generation method of Conditional Cycle Generative Adversarial Networks(CC-GAN).In this paper,we use the method of cyclic reconstruction to maintain facial identity features,generate pose loss to control the rotation of faces in different poses to the front view,and control the details of the image through feature loss.This paper proves through experiments on the FERET data set that the face frontalization method proposed in this paper can generate a front face image with clear details and can effectively improve the accuracy of face recognition.(3)Multi-pose face generation.Considering that the multi-pose face network model based on the general Encoder-Decoder structure pays more attention to the attitude problem in the face generation process.In order to generate a face image of the target pose,the decoder will lose some details.This paper proposes a Multitask Detail Compensated Generative Adversarial Networks(MDC-GAN)based framework to enhance the details of generating faces.We introduce an auxiliary network in the network to extract the facial features of the target pose,and deeply integrate the features of the input image to make up for the lack of details in the generation process.The multi-task learning mechanism is used to split the generation of multi-pose faces into three task stages: feature extraction,face synthesis and face reconstruction.The three networks are trained synchronously,and the performance of the model is improved through parameter sharing.Experiments on the sub-library of the FERET dataset prove that the network model proposed in this paper can generate multi-pose face images with clear details and textures,and can improve the accuracy of multi-pose face recognition.In summary,we propose to eliminate the influence of pose change on face recognition by using face frontalization and multi-pose face generation methods.Through experimental verification,the proposed method analyzes the experimental results from two aspects of visualization and objective indicators.The experimental results show that this method has a good effect on improving the accuracy of face recognition.