Video facial micro-expression and macro-expression detection method based on point-level weak supervision

By constructing a four-module detection framework and a point-level weakly supervised method trained with multiple loss functions, the problems of high annotation cost, low detection accuracy, and large pseudo-label noise in facial expression detection are solved. This method achieves accurate detection of micro-expressions and macro-expressions and is applicable to scenarios such as public safety, judicial evidence collection, and clinical diagnosis.

CN122244927APending Publication Date: 2026-06-19YANGTZE DELTA REGION INST OF UNIV OF ELECTRONICS SCI & TECH OF CHINE (HUZHOU)
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
YANGTZE DELTA REGION INST OF UNIV OF ELECTRONICS SCI & TECH OF CHINE (HUZHOU)
Filing Date
2026-05-09
Publication Date
2026-06-19

AI Technical Summary

Technical Problem

Existing facial expression detection methods suffer from high annotation costs, low detection accuracy, large pseudo-label noise, and insufficient feature learning, especially in micro-expression detection, making it difficult to achieve both low annotation costs and high detection accuracy.

Method used

We employ a point-level weakly supervised video face micro-expression and macro-expression detection method. By constructing a four-module detection framework, designing a multi-refined pseudo-label generation algorithm and a distribution-guided feature comparison learning module, and combining multiple loss functions for end-to-end training, we achieve high reliability of pseudo-labels and global feature representation capabilities.

Benefits of technology

It achieves a balance between low annotation costs and high detection accuracy, with a simple framework structure, and is suitable for scenarios requiring precise facial emotion analysis, such as public safety, judicial evidence collection, and clinical diagnosis, providing reliable technical support.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122244927A_ABST
    Figure CN122244927A_ABST
Patent Text Reader

Abstract

This invention discloses a video-based method for detecting micro-expressions and macro-expressions based on point-level weak supervision. The method includes: acquiring untrimmed face videos and preprocessing them into video segments; labeling each frame within the real expression range with point-level tags and aggregating them into video-level tags; extracting image and optical flow motion features and enhancing them to generate fusion features and attention scores; detecting the fusion features to generate temporal class activation maps to determine segment-specific probabilities; constructing a multi-instance learning loss based on the video-level tags to initially locate expression segments; modulating the attention scores, and generating a low-noise refined pseudo-label set through threshold filtering and probability secondary modulation; filtering and differentially sampling the fusion features based on the refined pseudo-label set; optimizing feature representation through contrastive learning to achieve intra-class aggregation and inter-class separation of features; and configuring end-to-end training using multiple losses as the framework for the above detection process. This invention can significantly improve the localization accuracy and category discrimination ability of micro-expressions and macro-expressions with low annotation costs.
Need to check novelty before this filing date? Find Prior Art