Abstract: Micro-expressions are fleeting facial expressions that expose a genuine emotions that a person tries to conceal. An existing multi-task middle-level feature learning is improved, and a multi-task middle-level feature individualization learning method for micro-expression recognition is proposed. The same kind of micro-expressions of the same person are removed for calculating its interclass nearest neighbors. the different kind of micro-expressions of the same person are retained and the [k] value is reduced for calculating its interclass nearest neighbors. A middle-level feature with more discriminative information is generated by individualization learning method. Experimental results on spontaneous micro-expression database CASME2 show that the proposed method has better recognition performance.

Key words: micro-expression recognition, individualization learning, middle-level feature

摘要： 微表情是一种短暂的面部表情，揭示了一个人试图隐藏的真实情感。对现有的一种多任务中级特征学习方法进行了改进，提出了一种多任务中级特征个性化学习方法用于微表情识别。对于每个低级特征，计算类内[k]最近邻时，去除同一人的同类微表情；计算类间[k]最近邻时，保留同一人的不同类表情，并减小[k]值。采用个性化学习方法，生成具有更多判别信息的中级特征。在微表情数据集CASME2上的实验表明，所提出的方法具有更好的识别性能。

关键词: 微表情识别, 个性化学习, 中级特征