PDA-AR: Progressive low-light facial expression recognition via multi-level alignment

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Low-light facial expression recognition (FER) is challenging due to noise, blur, and texture loss.
• Existing image enhancement methods may introduce artifacts, hindering FER accuracy.
• A robust method is needed for effective knowledge transfer from normal-light to low-light facial images.

Purpose of the Study:

• To propose a progressive framework (PDA-AR) for low-light FER.
• To enable knowledge transfer from normal-light to low-light images without explicit enhancement.
• To improve the accuracy and generalization of FER systems in challenging lighting.

Main Methods:

• Developed a progressive low-light facial expression recognition framework (PDA-AR).
• Implemented a Learning Sorting Module (LS) for brightness-based image sequencing.
• Utilized multi-level information alignment via Maximum Mean Discrepancy and Structural Similarity to bridge the domain gap.

Main Results:

• Achieved 87.81% recognition accuracy on low-light facial expression datasets.
• Demonstrated superior performance compared to existing mainstream methods.
• Validated the framework's generalization ability and practical value in low-light conditions.

Conclusions:

• The proposed PDA-AR framework effectively addresses low-light FER challenges.
• Knowledge transfer without enhancement is feasible and beneficial for low-light FER.
• PDA-AR offers a promising solution for robust facial expression recognition in adverse lighting.