Abstract:

Micro-expression recognition is still a challenging task due to its short duration of less than 0.5?s, involuntariness, and low intensity. The hierarchical spatio-temporal feature descriptor is improved, and a new fine-grained hierarchical spatio-temporal feature micro-expression recognition method is proposed. Firstly, the spatio-temporal features of each level in the micro-expression video clip are extracted, and the projection matrix is used to establish the relationship between the spatio-temporal features and the micro-expressions, and then the regions that contribute to the recognition task are selected. Then it counts the layer with the overall maximum contribution, and performs the intersection operation between the selected blocks at the lower layer and the selected blocks in the previous layer to achieve the purpose of removing the spatial redundancy of the hierarchical spatio-temporal features and improving the discrimination of the micro-expression features. Experiments on CASME[Ⅱ] show that the proposed method can fine-grain the micro-expression area and obtain better recognition results.

Key words: micro-expression recognition, hierarchical spatio-temporal feature, fine-grained

摘要：

由于微表情持续时间小于0.5?s、非自愿性和低强度等特点，微表情识别仍然是具有挑战性的任务。对分层时空特征描述符进行改进，提出一种新的细粒度分层时空特征的微表情识别方法。提取微表情视频片段中的各层次时空特征，利用投影矩阵建立时空特征和微表情之间的联系，进而选择对识别任务有贡献的区域。然后统计具有整体最大贡献度的层次，将该层次下选中的区域块和前一层选中的区域块进行交集操作，达到去除分层时空特征的空间冗余性和提升微表情特征区分度的目的。在CASME[Ⅱ]上的实验表明，提出的方法能够细粒度化微表情发生区域，获得了更好的识别结果。

关键词: 微表情识别, 分层时空特征, 细粒度