Deep 3D-LBP: CNN-based fusion of shape modeling and texture descriptors for accurate face recognition

The key challenge of face recognition is to develop an effective feature representations for reducing intrapersonal variations while enlarging interpersonal differences. In this paper, we show that the face recognition accuracy may be enhanced with the combination of a 3D model-based alignment, and an LBP descriptor constructed on the 3D mesh. First, 3D face data are reconstructed from 2D images that aim to normalize the input image. Then, shape and texture features on the mesh are extracted using the mesh local binary patterns: mesh-LBP. With the use of the extracted 3D features and a simple CNN architecture, much higher accuracy rates can be achieved. We achieve the accuracy of 99.59% on the widely used labeled Faces in the Wild dataset. On YouTube Faces dataset, the proposed method achieves 94.97%, despite using a small training dataset.