基于HMM的单样本可变光照、姿态人脸识别

Pose and Illumination Invariant Face Recognition Based on HMM with One Sample Per Person

提出了一种基于HMM的单样本可变光照、姿态人脸识别算法.该算法首先利用人工配准的训练集对单张正面人脸输入图像与Candide3模型进行自动配准,在配准的基础上重建特定人脸三维模型.对重建模型进行各种角度的旋转可得到姿态不同的数字人脸,然后利用球面谐波基图像调整数字人脸的光照系数可产生光照不同的数字人脸.将产生的光照、姿态不同的数字人脸同原始样本图像一起作为训练数据,为每个用户建立其独立的人脸隐马尔可夫模型.将所提算法对现有人脸库进行识别,并与基于光照补偿和姿态校正的识别方法进行比较.结果显示,该算法能有效避免光照补偿、姿态校正方法因对某些光照、姿态校正不理想而造成的识别率低的情况,能更好地适应光照、姿态不同条件下的人脸识别.

In this paper a novel pose and illumination invariant face recognition algorithm that based on HMM with one sample per person is proposed. Firstly, by learning the train sets that has been fitted with Candide3 model, fitting algorithm automatically fitted the frontal facial input image with Candide3 to get the 3D shape. Then the specifically 3D face model is reconstructed by synthesizing texture to the fitted 3D shape. The new images under different pose can be generated by transform the 3D face model. Decompose the new images to 9 harmonic images＇ linear combination. By changing the 9 coefficients the new images under different illumination condition can be generated. All of these new images under different pose and different illumination condition make up the train set of individual face hmm. Experimental results show that this method can effectively avoid the recognition rate bring down caused by the pose and illumination normalization is not effective sometimes and can be better fitting the pose and illumination invariant face recognition.