Face Frontalization Using an Appearance-Flow-Based Convolutional Neural Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biometrics

Background:

• Facial pose variation significantly challenges face recognition (FR).
• Current CNN methods synthesize frontal faces in color space, often losing fine facial textures due to non-linear learning.
• Pixel changes in face frontalization are primarily due to geometric transformations.

Purpose of the Study:

• To develop a novel method for face frontalization that preserves fine facial textures.
• To improve the accuracy and robustness of face recognition systems under varying poses.
• To address the limitations of color-space-based synthesis in existing methods.

Main Methods:

• Proposing an appearance-flow-based convolutional neural network (A3F-CNN) for spatial domain face frontalization.
• Learning dense correspondence between non-frontal and frontal faces to explicitly "move" pixels.
• Employing an appearance-flow-guided learning strategy, generative adversarial network loss, and a face mirroring technique.

Main Results:

• A3F-CNN successfully synthesizes photorealistic frontal faces, preserving fine facial textures.
• The method outperforms existing techniques in both controlled and uncontrolled lighting conditions.
• Achieved competitive performance in pose-invariant face recognition on benchmark datasets (Multi-PIE, LFW, IJB-A).

Conclusions:

• Spatial domain face frontalization is more effective than color domain synthesis for preserving facial details.
• The proposed A3F-CNN method offers a robust solution for generating high-quality frontal faces from varied poses.
• This approach significantly enhances face recognition performance, particularly in challenging pose variations.