A facial depression recognition method based on hybrid multi-head cross attention network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Convolutional Neural Networks (CNNs) show promise in depression analysis but struggle with learning long-range facial feature dependencies due to spatial locality.
• Single attention mechanisms limit models' ability to focus on multiple crucial facial regions simultaneously for depression detection.

Purpose of the Study:

• To develop an advanced deep learning framework, the Hybrid Multi-head Cross Attention Network (HMHN), to overcome limitations in current CNN-based depression recognition.
• To improve the accuracy and sensitivity of video-based depression analysis by effectively integrating information from diverse facial areas.

Main Methods:

• The HMHN framework employs a two-stage approach: Grid-Wise Attention (GWA) and Deep Feature Fusion (DFF) for low-level feature extraction.
• The second stage utilizes a Multi-head Cross Attention block (MAB) and Attention Fusion block (AFB) to encode high-order interactions among local features for global representation.

Main Results:

• Experiments on the AVEC2013 and AVEC2014 depression datasets demonstrated the HMHN's effectiveness.
• The proposed method achieved superior performance, outperforming most state-of-the-art video-based depression recognition approaches with reported RMSE and MAE values.

Conclusions:

• The HMHN model successfully captures higher-order interactions between depression features from multiple facial regions, significantly reducing recognition errors.
• This deep learning approach holds substantial potential for clinical applications in depression assessment and diagnosis.