Stacked attention hourglass network based robust facial landmark detection

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning has advanced facial landmark detection (FLD).
• Existing FLD algorithms struggle with accuracy and robustness due to diverse expressions, poses, and occlusions.
• Outliers significantly degrade performance in facial landmark detection.

Purpose of the Study:

• To develop a more discriminative representation for facial landmark detection.
• To enhance the robustness of facial landmark detection algorithms against various challenging conditions.
• To reduce the negative impact of outliers on facial landmark detection accuracy.

Main Methods:

• A stacked attention hourglass network (SAHN) is proposed for FLD.
• Introduced a spatial attention residual (SAR) unit to emphasize relevant facial areas and extract multi-scale features.
• Incorporated a channel attention branch (CAB) to guide feature extraction in subsequent network levels.
• Developed a variable robustness (VR) loss function for adaptive training and improved model resilience.

Main Results:

• The proposed SAHN, with fewer parameters than traditional models, demonstrated superior performance.
• Experimental results on 300W, WFLW, and COFW datasets confirmed the method's effectiveness.
• The integration of SAR and CAB modules led to improved feature representation and robustness.

Conclusions:

• The novel SAHN with attention mechanisms and VR loss significantly improves facial landmark detection accuracy and robustness.
• The proposed method offers a more efficient and effective solution for FLD in unconstrained environments.
• This work advances the state-of-the-art in facial landmark detection by addressing key challenges in real-world scenarios.