基于渐进式稀疏堆叠自动编码器的姿态不变的人脸识别方法

Pose-Invariant Face Recognition (PIFR) is a challenging problem since it is difficult to learn geometrically invariant feature representation for large pose face images. On the one hand, the pose variations of faces is a highly non-linear process. On the other hand, pose variations will lead to the loss of a large proportion of discriminative face information. And in the case of insufficient training data, there are risks of over fitting in the complex models. In order to solve these problems, this paper proposes a pose-invariant face recognition method based on sparse multiple auto-encoders, named as Sparse Stacked Progressive Multiple Auto-Encoders (Sparse SPMAE): (1) The non-linear transform process from the pose face images to frontal ones is divided into several nearly linear process. By stacking multiple auto-encoders with adaptive hidden layers, the face normalization is implemented step by step in a reverse and progressive manner. (2) It introduces the sparse constraints in multiple auto-encoders to learn the effectively discriminative information and reduce the potentials of over fitting in the Sparse SPMAE models. As a result, Sparse SPMAE obtains good accuracy on three classic datasets.