Multi-Stage Statistical Texture-Guided GAN for Tilted Face Frontalization.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biometrics

Background:

• Existing face recognition methods struggle with high-pitch angle variations common in surveillance.
• Self-occlusion in tilted faces significantly degrades feature extraction for accurate recognition.

Purpose of the Study:

• To develop a robust method for high-pitch angle face recognition by addressing the challenge of face frontalization.
• To improve the accuracy and reliability of face recognition systems under extreme pose variations.

Main Methods:

• A Statistical Texture-guided Generative Adversarial Network (STG-GAN) is proposed for tilted face frontalization.
• The STG-GAN employs a face encoder, statistical texture modeling, and a pose-guided decoder.
• A divide-and-conquer strategy with multi-stage progressive synthesis and specialized losses (content, identity, adversarial, pose contrastive) is utilized.

Main Results:

• The STG-GAN successfully reshapes high-pitch angle faces into approximate frontal views.
• The method demonstrates superior performance in qualitative and quantitative experiments across multiple face datasets.
• Pose-invariant feature extraction and accurate facial component generation are achieved.

Conclusions:

• The proposed STG-GAN offers a significant advancement in pose-invariant face recognition, particularly for high-pitch angles.
• This approach enhances the reliability of face recognition in real-world surveillance applications.
• The statistical texture-guided frontalization effectively handles self-occlusion and pose variations.