3D Face Recognition Based On Convolution Neural Network

Face recognition as a hot research direction in the field of computer vision and artificial intelligence has got much attention. Face recognition technology has experienced decades of development, but since the face in the real scene is too complex, now face recognition technology is still difficult to apply in actual scenario. This thesis seize the 3D depth information which is often ignored but very important, and use the deep learning method to analysis the importance of 3D information in face recognition. This thesis also puts forward kinds of 3D and 2D information fusion strategies.This thesis selects the convolution neural network as the basic research method, and build a CNN structure with eight layers based on the characteristics of the experiment database. This network adapt two convolution layers and two pooling layers, one local connection layer, two full connection layers and one regression layer. This network not only has a small number of parameters, but also can achieve high recognition accuracy.The small number of parameters can avoid over fitting during training and can reduce the network’s dependency of data. The experimental results show that this network can get a good enough recognition result with very little training data.In order to verify the effectiveness of the 3D face information, this thesis takes contrast experiments using the eight layers convolution neural network. The experimental results show that using 3D face data alone is more efficient than using 2D face data alone.Then this thesis propose three different kinds of 3D and 2D fusion strategies, respectively are RGB-D, Gray-D and RGB-D-W. The final experimental results show that,without losing any information, the RGB-D strategy get the highest recognition accuracy.Adopting deep learning method which can extract nonlinear and advance features, this thesis studies the importance of 3D information in face recognition field. This thesis propose a convolution neural network structure, and using the network to analysis and compare the 3D and 2D information, then find a good way to fuse 3D and 2D information.