AAN-Face: Attention Augmented Networks for Face Recognition

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional neural networks (CNNs) excel at feature extraction for face recognition.
• CNNs struggle with imbalanced datasets, leading to poor generalization, especially for under-represented classes like occluded or profile faces.
• Recognizing masked faces, critical during the COVID-19 pandemic, highlights the challenge of data imbalance and occlusions.

Purpose of the Study:

• To develop an Attention Augmented Network (AAN-Face) to improve face recognition performance on imbalanced datasets.
• To enhance model robustness against occlusions and pose variations.
• To specifically address the challenge of recognizing masked faces.

Main Methods:

• Proposed an attention erasing (AE) scheme to randomly remove units in attention maps, preparing models for occlusions and pose variations.
• Introduced an attention center loss (ACL) to learn centers for attention maps, ensuring focus on consistent facial parts and suppressing noise.
• Integrated AE and ACL into the AAN-Face model, encouraging the localization of diverse and complementary facial parts.

Main Results:

• AAN-Face demonstrated superior performance compared to state-of-the-art methods on various datasets.
• Significant improvements were observed in recognizing minority classes, including masked faces.
• The proposed methods effectively emphasized discriminative facial regions while suppressing irrelevant ones.

Conclusions:

• AAN-Face effectively handles imbalanced data distributions in face recognition tasks.
• The combination of attention erasing and attention center loss enhances model generalization and robustness.
• The approach shows significant promise for real-world applications, particularly in scenarios with occlusions like masked faces.