ArecaNet: Robust Facial Emotion Recognition via Assembled Residual Enhanced Cross-Attention Networks for...

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Facial emotion recognition (FER) is crucial for human-computer interaction (HCI).
• Traditional methods and single deep learning networks (CNNs, ViTs) face limitations in accuracy and information loss.
• Overfitting and single-network dependency hinder performance in existing FER systems.

Purpose of the Study:

• To introduce ArecaNet, an assembled residual enhanced cross-attention network for superior FER.
• To overcome limitations of single networks and conventional ensemble methods in FER.
• To achieve state-of-the-art accuracy in facial emotion recognition.

Main Methods:

• Developed ArecaNet, integrating channel, spatial, and landmark features.
• Employed SCSESResNet for feature extraction and specialized sub-networks for landmark data.
• Utilized iterative residual enhanced cross-attention for seamless feature fusion and minimized information loss.

Main Results:

• ArecaNet achieved 97.0% accuracy on FER-2013 and 97.8% on RAF-DB.
• Outperformed the previous state-of-the-art method, PAtt-Lite, by 4.5% on FER-2013 and 2.75% on RAF-DB.
• Established a new state-of-the-art accuracy on both public databases.

Conclusions:

• ArecaNet effectively integrates diverse facial features using an iterative attention mechanism.
• The proposed network architecture overcomes information loss and single-network dependency issues.
• ArecaNet demonstrates significant advancements in facial emotion recognition accuracy and robustness.