Research And Implementation Of Co-Learning For Face Reconstruction And Face Recognition

Monocular face reconstruction and face recognition tasks are hot topics in the field of computer vision in recent years.The task of combining these two tasks mainly aims to solve the problem of recognizability of face reconstruction.This paper firstly studies and implements the classic face reconstruction algorithm based on single face image,and then combines face learning and overcome the problem that there is no existing multi-labels for face reconstruction and recognition in academic circles.The reconstructed face shape achieves high performance on the face recognition test set.Firstly,in the classic face reconstruction algorithm based on single image,this paper uses 3DMM(3D Morphable Model)as the basic geometric model of the face,and then parameterize the camera.The open source 300W-3D data set is used as a training set.And this task is trained on 64-layers CNN,and the 3DMM coefficients and camera parameters are regressed with respect to the input face image.Finally,the parameters are reconstructed based on these parameters.Based on the basic single image reconstruction algorithm,this paper combines the face recognition algorithm to cluster the shape coefficients of 3DMM,which makes the 3DMM coefficients have high recognizability.In order to achieve this goal,this paper selects several different data sets and manually generates tags of some data sets to meet the training requirements.Because the convergence direction of face recognition and reconstruction is completely different,this paper designs the CNN network structure in a targeted manner,and splits the training steps into multiple stages.At each stage,one target module will be pre-trained while other modules are locked.Finally,all modules in the network are opened for training.At the same time,the loss function designed in this paper combines the tasks of face reconstruction and recognition,and considers the problem that different training images have different labels,so that the whole network can finally converge.Finally,the joint learning algorithm of face reconstruction and face recognition designed in this paper achieves the accuracy comparable to PRNet and 3DDFA on the AFLW2000-3D dataset,and achieves an RMSE score of 1.85 on the MICC face reconstruction dataset.On some popular face identification databases CFP-FP,AgeDB,LFW and YTF,this paper reached 95.10%,90.30%,99.32% and 96.66% respectively,which verified the effectiveness of the proposed algorithm.