关键词: 反投影概率图, 多状态部件模型, 卷积神经网络, 形状参数更新, 分层人脸配准

Abstract: The range of expression variety is discretized to be multi-state component model, which is used to describe nonlinear variety of a face. The hierarchical face alignment is achieved based on the cascade convolutional neural networks by introducing multi direction local gradient information and constructing the corresponding backprojection probability map to improve appearance representation of traditional grey images. The face shape initialization is realized and the current component state is predicted, according to the whole face and different regions. The face shape parameters are regressed to achieve final fine alignment, according to the chosen face model under correct state. The experiments performed on database demonstrate that the proposed method improves the robustness of face alignment under exaggerated expressions and correct rate, and convergence rate are increased under the considerable computational complexity.

Key words: backprojection probability map, multi-state component model, Convolutional Neural Network, shape parameters update, hierarchical face alignment