Single Image Based Three-Dimensional Face Reconstruction And Recognition

3D face perception is a hot research topic in computer vision and pattern recognition,which has received extensive attention.The dense correspondence,reconstruction and recognition of 3D face are closely related.3D face dense correspondence builds a point-to-point correspondence between faces,it can bring convenience to the research of 3D face.3D face reconstruction restores face shape,which has broad application in animation,face recognition and etc.3D face recognition can solve the problem of pose and illumination in 2D face recognition,it is more robust when we apply the face recognition technology in natural scenes.The main work of this paper is as follows.Firstly,a 3D face dense correspondence based on improved Non-rigid iterative closest point(NICP)is proposed.The NICP-based methods deform a face template to be close with target face,but these algorithms preserve the topology by sacrificing the precision.A constraint of topology was proposed so it not need to greatly reduce the weight of the distance loss,hence the precision can be improved.For the problem that there are some points with big error in the central of the face,a weight adaptive strategy is used to improve the precision.Experimental results show that the proposed algorithm can improve the precision while maintaining the face topology.Secondly,a 3D face reconstruction method based on sample-selecting is proposed.The 3D morphable model and its improved methods are based on the linear-class assumption,a 3D face is in the subspace expand by its similar faces.Representing a target face with all faces in the database has not consider the difference between target face and its non-similar samples,it will cause a large deviation.The classification algorithm was used to determine the gender and expression of the input image,the corresponding 3D face subset was selected according to the gender and expression,then the sparse representation theory was used in the subset for selecting the similar samples,next the morphable model was constructed with the selected face,finally matching the model with feature points in the target face.Experimental results show that the reconstructed faces have strong realism and high precision,and the algorithm has the ability to reconstruct the expressive face.Finally,a 3D face recognition method based on convolutional neural network(CNN)is proposed.The depth map used for projecting 3D data to 2D plane.Aiming at the problem that the depth information is greatly lost when obtaining the depth map,a cylindrical depth map with local information enhancing is designed.Then the improved depth map and texture image are combined to get the 3D+2D feature map.Finally,a CNN model is trained based squeeze-net,and squeeze-net was cropped in order to decrease recognition time.Experiments show that the proposed algorithm has high recognition accuracy and good robustness.