3D Face Recognition Based on Twin Neural Network Combining Deep Map and Texture

Massive amount of training samples is a challenge for 3D face recognition using deep learning frame. This paper shows a method that uses deep twin neural network for 3D face recognition by blending face 3D depth and 2D texture. First, a depth map is generated. In order to repair holes in the 3D face model with low complexity, we map those 3D hole points into 2D plane, and then reverse them back to 3D space by the least square rule. Second, a convolution kernel model with two layer channels is used to fuse face image and depth image. Finally, after sample pairs are generated, 3D face recognition is performed by convolutional twin neural network. The experimental results on CASIA-3D dataset show that, compared with the classical CNN method, the recognition accuracy of our method increases about 2.85%. And in the case of using small training sets, the recognition rate of our method is about 4% higher.