.png){kind=icon}
.png){kind=icon}
Return
Micro-expression recognition using quantitative feature validity and pseudo-emotion feature construction
.png){kind=icon}
PRE
AI
F
L
DOI：10.1007/s00530-026-02569-3.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Due to the limited data on micro-expressions, providing richer emotion information by constructing multifeature forms is one of the commonly adopted approaches for micro-expression recognition. However, this small sample problem leads to the fact that the extracted emotion features contain various interference information. It also lacks some adaptability to the problem of feature selection and fusion in multifeature settings, which limits the development of micro-expression recognition. Based on this, we propose a micro-expression recognition algorithm that leverages quantitative feature validity and pseudo-emotion feature construction to address interference and fusion selection issues. For the interference problem in the features, we proposed the concept of pseudo-emotion for the first time. We extracted pseudo-emotion features using a convolutional network trained on onset frames with emotion labels, effectively removing interference information from the features and improving their representational capacity. For the problem of how to evaluate the validity of features and then adaptive feature selection and fusion under multifeature form, we first propose a quantitative feature validity metric, which can quantize the extracted feature vectors based on the classification space of the samples, and then realize the efficient and accurate adaptive fusion selection based on the quantitative metrics. Finally, micro-expression recognition is performed using deep networks. The validity and rationality of the algorithm are verified through a comprehensive experimental analysis of public datasets. It also significantly outperforms other state-of-the-art algorithms. In the three classification experiments, the highest metric reached 0.97.
Keywords:
Convolutional networks
Feature validity
Micro-expression recognition
Pseudo-emotion
Journal
IF:
3.1
Papers:
2.7K
Citations:
2.7K
Organization
Cited Papers