Abstract: For a long time, the issue of limited samples has been a major hindrance to the development of micro-expression analysis, and this limitation primarily stems from the inherent difficulty in annotating micro-expression data. In this research, we aim to address this challenge by leveraging facial electromyography as a technical approach and propose three solutions for micro-expression data annotation: automatic annotation, semi-automatic annotation, and unsupervised annotation. Specifically, we first present an automatic micro-expression annotation system based on distal facial electromyography. Second, we propose a semi-automatic annotation scheme for micro-expression onset and offset frames based on single-frame annotation. Finally, for unsupervised annotation, we introduce a cross-modal self-supervised learning algorithm based on electromyographic signals. Additionally, this research endeavors to explore the temporal and intensity characteristics of micro-expressions using the electromyography modality.

长久以来,微表情的小样本问题始终制约着微表情分析的发展,而小样本问题归根到底是因为微表情的数据标注十分困难。本研究希望借助面部肌电作为技术手段,从微表情数据自动标注、半自动标注和无标注三个方面各提出一套解决方案。对于自动标注,提出基于面部远端肌电的微表情自动标注方案;对于半自动标注,提出基于单帧标注的微表情起止帧自动标注;对于无标注,提出了基于肌电信号的跨模态自监督学习算法。同时,本研究还希望借助肌电模态,对微表情的呈现时间和幅度等机理特征进行拓展研究。

国家自然科学基金项目(62276252, U19B2032, 62106256)王甦菁