Research On 3D Face Recognition Technology Based On Deep Learning

Face recognition technology,one of the biometric recognition technologies,has attracted widely attention from researchers due to its advantages of naturalness,non-contact,uniqueness and initiative.With the rapid development of science and technology,face recognition technology has also made great progress.Two-dimensional(2D)face recognition technology has been maturing and is widely used in our lives.The essence of 2D face image is a simple projection of the three-dimensional(3D)face image into a 2D space,which lacks such spatial information of the 3D face image as face information of curved surface,of curvature,and of spatial structure.Therefore,2D face recognition is inevitably affected by factors like illumination,pose and expression.Compared with the 2D face image,the3 D face image is much more in line with the characteristics of human vision and has the spatial stereo information of the real face.Thus,the research on the 3D face recognition technology has become a foucus area in the face recognition field.In view of the poor robustness of the two-dimensional face recognition method to changes in illumination,posture,and expression,based on the Texas 3DFRD three-dimensional face database,this thesis conducts research on three-dimensional face recognition technology based on the deep learning.The specific work of the thesis is as follows:Firstly,this thesis takes a research about the classic classification network VGG(Visual Geometry Group,VGG)convolutional neural network in deep learning.Combining different features at different scales,a cross layer VGG convolutional neural network(Cross-layer Visual Geometry Group,CLVGG)has been proposed in this thesis.The convolutional neural network of classic classification algorithm,VGG,is improved,including network structure and parameters.In terms of network structure,a cross-layer connection between the first convolution layer and the pooling layer is proposed in this thesis,which can add different scale features for high recognition accuracy.In terms of network parameters,the number of convolution kernels is increased in the last convolutional layer,and one of the fully connected layers is removed,which can reduce the overall parameters of the network model.Then,on the basis of the Texas 3DFRD three-dimensional face database,therecognition accuracy and operating efficiency of the VGG network and the CLVGG network are compared and analyzed.The experimental has shown that the the proposed CLVGG has high performace both in recognition accuracy and operating efficiency.Then,on the basis of the 3D face database of the Texas 3DFRD,2D,3D face recognition and multimodal fusion face recognition has been accomplished by the convolutional neural network of improved VGG.In the experiments the robustness to lighting and expression has been analyzed.The experimental results show that the 3D face information can effectively improve the robustness to face recognition under the influence of lighting and expression.At last,based on the Kinect V2 depth camera and graphical user interface application development framework Qt,a three-dimensional face recognition system is designed and implented.The algorithm proposed in this thesis is applied to the system,and on the basis of the self-built three-dimensional face database,the correctness,feasibility and effectiveness of the system are verified.