基于跨域异质度量迁移学习的心理应激微反应自动分析方法研究

As fraud detection clues, psychological stress micro-expression is of great value in such areas as national security, investigative trial and human-computer interaction. Automatic analysis technology based on image recognition puts particular emphasis on the narrow sense of "micro-expression" level, and also it suffers from lacking labeled training samples since labeling a large number of samples is a time consuming and labor intensive process. It still remains a critical problem how to establish the reliable classification model for target-domain data (say "micro-expression") and rely upon labeled source-domain data (say "micro-semantic"). Therefore, this project will focus on methods of automatically analyzing psychological stress micro-expression based on cross-domain heterogeneous metric transfer learning. Three methods on the cross-domain heterogeneous metric transfer learning theory are to be proposed: (1) sparse coupled metric learning from the feature selection perspective; (2) domain adaptation metric learning with source domain targetized from the cross-domain reconstruction view; (3) from the regression analysis view, i.e., heterogeneous metric transfer learning based on regression analysis method. Supported by this project, the proposed methods are potential solutions to the problems like feature matching between "micro-expression" and "micro-semantic", making up for the defects that influence of subjective factors due to that fact that psychologists rely on trained professionals to judge the psychological stress micro-expression. The derived research results can be extended and applied in various cross-domain data feature matching, which will benefit the correlated engineering practice in both theoretical and methodological aspects.

心理应激微反应作为欺骗检测线索，在国家安全、侦查审讯、人机交互等领域有重要价值。基于图像识别的自动分析技术侧重在狭义的“微表情”方面，并且受困于有标签的训练样本少的问题，标注大量的样本耗时耗力。如何依靠源领域（“微语义”）数据，建立可靠的模型用于目标领域（“微表情”）数据分类尚待研究。本项目将开展基于跨域异质度量迁移学习的心理应激微反应自动分析方法研究：（1）从特征选择的角度出发，提出稀疏耦合度量学习方法；（2）从跨域重建的角度出发，提出源域目标化的领域适应度量学习方法；（3）从回归分析的角度出发，提出回归分析的异质度量迁移学习方法。本项目拟研究的方法有利于解决“微表情”和“微语义”跨域特征匹配的问题，弥补心理学家依靠专业训练的人员来判断而受到的主观因素影响的缺陷。研究成果易推广到各种跨域数据的特征匹配问题上，为相关工程实践提供一定的理论支持和科学方法。

在本项目的资助下，我们在核心刊物和会议上共发表了学术论文19篇，包括IEEE Transactions on Image Processing, IEEE Transactions on Circuits and Systems for Video Technology, Pattern Recognition 等期刊论文，申请了24项国家发明专利，18项已经授权。心理应激微反应作为欺骗检测线索，在国家安全、侦查审讯、人机交互等领域有重要价值。基于图像识别的自动分析技术侧重在狭义的“微表情”方面，并且受困于有标签的训练样本少的问题，标注大量的样本耗时耗力。如何依靠源领域（“微语义”）数据，建立可靠的模型用于目标领域（“微表情”）数据分类尚待研究，也是本项目主要解决的问题。本项目开展了基于跨域异质度量迁移学习的心理应激微反应自动分析方法研究：（1）从特征选择的角度出发，提出了基于更新标记向量的耦合源域目标化模型；（2）从跨域重建的角度出发，提出了基于奇异值分解的领域适应度量学习模型；（3）从共同表示学习角度出发，提出了基于耦合度量学习的模型；（4）从回归分析的角度出发，提出了迁移支持向量机模型。此外，依托本项目，我们还建立了一个微表情数据库。本项目研究的方法有利于解决“微表情”和“微语义”跨域特征匹配的问题，弥补心理学家依靠专业训练的人员来判断而受到的主观因素影响的缺陷。研究成果可以推广到各种跨域数据的特征匹配问题上，为相关工程实践提供一定的理论支持和科学方法。