基于LLR算法的多姿态人脸识别

Multi-pose face recognition based on LLR algorithm

多姿态人脸识别是模式识别领域的难点之一,针对该问题提出的很多效果较好的算法都有其局限性,不能适应人脸状态多变的特征。在局部线性回归算法(LLR)的基础上,对人脸进行归一化校正,并通过引入一个局部常量因子,对不同水平旋转角度的侧脸进行姿态估计得到其正脸。改进后的LLR算法对人脸识别率有较大改善,这表明采用局部常量化和线性化分析,可以较好地弥补侧脸到正脸变换的非线性信息丢失。

Multi-pose face recognition is one of the most difficult problems in pattern recognition field. In order to solve this problem, put forward many algorithms, but almost none of them could adapt to the changing characteristics of the human faces. In this thesis, normalized both side faces and their corresponding frontal faces based on LLR algorithm, and by introdu- cing a new local constant factor to the traditional LLR, obtained frontal faces from side faces in different poses. The improved LLR can perform better face recognition, which shows that with the help of the local constant factor and local linearization, the nonlinear information loss of the transformation from side faces to frontal faces can be compensated effectively.