Confidence Preserving Machine for Facial Action Unit Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Biomedical Signal Processing

Background:

• Automated facial action unit (AU) detection is crucial for facial expression analysis but faces challenges from inter-personal variability, pose variations, and low-intensity AUs.
• These challenges lead to 'hard samples' that impede accurate detection.

Purpose of the Study:

• To propose a novel framework, the Confidence Preserving Machine (CPM), to address the issue of hard samples in facial AU detection.
• To improve the accuracy and robustness of automated facial expression analysis.

Main Methods:

• Developed a two-stage learning framework (CPM) employing an 'easy-to-hard' strategy with multiple classifiers.
• Introduced a quasi-semi-supervised (QSS) learning strategy utilizing spatio-temporal smoothness for person-specific classifier training.
• Proposed extensions: iterative CPM for sample augmentation and kernel CPM for enhanced nonlinearity.

Main Results:

• CPM models demonstrated significant benefits over baseline and state-of-the-art semi-supervised and transfer learning methods.
• Experiments on four spontaneous datasets (GFT, BP4D, DISFA, RU-FACS) validated the effectiveness of the proposed CPM approaches.
• The framework successfully improved detection performance by effectively handling challenging facial samples.

Conclusions:

• The Confidence Preserving Machine (CPM) offers a robust solution for accurate facial action unit detection, particularly in the presence of challenging samples.
• The proposed QSS learning strategy and CPM extensions provide valuable advancements in automated facial expression analysis.