DCAlexNet: Deep coupled AlexNet for micro facial expression recognition based on double face images

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Facial Micro-Expression Recognition (FER) faces challenges from individual emotional variations and complex feature extraction.
• The role of apex frames in FER is not fully understood, and low-resolution images hinder performance.
• Existing super-resolution and CNN methods offer limited improvements for FER.

Purpose of the Study:

• To propose a novel deep coupled AlexNet (DCAlexNet) model for enhanced FER.
• To investigate the integration of global and local facial information for improved micro-expression detection.
• To address performance degradation caused by low-resolution facial images.

Main Methods:

• Developed a deep coupled AlexNet (DCAlexNet) with a trunk network for multi-resolution feature extraction.
• Implemented a branch network for resolution-specific mapping between high-resolution (HR) and low-resolution (LR) images.
• Integrated global and local facial information while filtering irrelevant facial regions.

Main Results:

• Achieved superior performance on FER2013, BU-3DFE, and Oulu-CASIA datasets.
• Attained 98.3% accuracy on FER2013, 97.2% on BU-3DFE, and 96% on Oulu-CASIA.
• Demonstrated improvements in RMSE, RAE, and processing times.

Conclusions:

• DCAlexNet effectively enhances micro-expression recognition, particularly for low-resolution images.
• The model's ability to integrate multi-resolution data and filter regions improves FER accuracy.
• This approach offers a significant advancement in automated facial micro-expression analysis.