基于自适应鲁棒在线度量学习的面部表情识别

Facial Expression Recognition Based on Adaptive Robust Online Metric Learning Algorithm

针对自然面部表情识别中的噪声标记问题,提出了一种自适应鲁棒在线度量学习方法。首先,学习新的度量空间以增加不同面部表情的判别性;然后,定义敏感度和特异性来表征每个注释器;最后,引入表示真实类标签的潜在变量,在期望最大化架构中迭代求解距离度量和注释器的可靠性。在MFP和AR人脸数据库上的实验结果表明,相比其他几种较新的方法,本方法在自然表情识别方面能获得更高的识别精度,高兴表情识别率可高达99.7%,并且在一定程度上降低了计算开销。

For the problem ofnoise tag in spontaneous facial expression recognition,an adaptive robust online metric learning method is proposed.Firstly,a new metric space is learned to increase the discrimination of different facial expressions.Then,the sensitivity and specificity is defined to represent each annotator.The latent variables representing real class label are introduced,and iteration is used to solve the distance measure and the reliability of the annotator in expectation maximization architecture.Experimental results on MFP and AR face databases show that proposed method has higher recognition accuracy than several other advanced methods,its accuracy on happy expression can achieve 99.7％,and it has partly reduced the computation overhead.