DotFAN: A Domain-Transferred Face Augmentation Net.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Network (CNN) performance in face recognition heavily depends on diverse, labeled training data.
• Acquiring large datasets with variations in pose and illumination for face recognition is costly and difficult.
• Within-class diversity in face image datasets is a critical challenge for robust face recognition systems.

Purpose of the Study:

• To introduce DotFAN, a novel 3D model-assisted domain-transferred face augmentation network.
• To address the challenge of limited within-class diversity in face recognition training datasets.
• To generate realistic face image variants that improve the performance of face recognition models.

Main Methods:

• DotFAN extends the StarGAN architecture, incorporating a Face Expert Model (FEM) and a Face Shape Regressor (FSR).
• The Face Shape Regressor (FSR) extracts facial attributes, while the Face Expert Model (FEM) captures identity information.
• This architecture enables separate learning of facial feature codes, allowing attribute manipulation while preserving identity.

Main Results:

• DotFAN successfully generates diverse face image variants from input faces.
• The generated images augment existing small face datasets, significantly increasing within-class diversity.
• Face recognition models trained on DotFAN-augmented datasets show improved performance.

Conclusions:

• DotFAN effectively addresses the scarcity of diverse training data in face recognition.
• The proposed method enhances within-class diversity, leading to more robust face recognition models.
• DotFAN offers a practical solution for improving face recognition system accuracy through data augmentation.