Micro-expressions are spontaneous emotions appearing on a face that is hard to conceal and thus making them different from normal facial expressions both in duration and subtlety. This paper investigates a challenging issue in micro-expression, where not all facial regions contribute equally to effective representation. Consequently, we proposed a multi-scale active patches fusion-based spatiotemporal LBP-TOP descriptor that considers the active contributions for different region area in faces. For the feature procedure, we exploit the average value of all patches under each scale to obtain the threshold that selectively fuses the local and global features. On the other hand, an improved weighted sparse representation based dual augmented Lagrange multiplier is adopted for the classification to remit the problem of sparse coefficients obtained by the traditional sparse representation algorithm. We conduct comprehensive experiments on CASME II and SAMM datasets and the accuracies respectively reach 77.30% and 58.82% using LOSO cross-validation.

微表情是出现在难以掩饰的脸上的自发情绪，因此使它们在持续时间和微妙程度上与正常面部表情不同。本文研究了微观表达中一个具有挑战性的问题，在该问题中，并非所有面部区域都同样有效地代表了面部表情。因此，我们提出了一种基于多尺度主动补丁融合的时空LBP-TOP描述符，该描述符考虑了面部不同区域面积的主动贡献。对于特征过程，我们利用每个尺度下所有补丁的平均值来获得阈值，以选择性地融合局部特征和全局特征。另一方面，采用改进的基于加权稀疏表示的对偶增强拉格朗日乘数进行分类，以解决传统稀疏表示算法获得的稀疏系数问题。我们对CASME II和SAMM数据集进行了全面的实验，使用LOSO交叉验证的准确性分别达到77.30％和58.82％。