Adversarial Cross-Spectral Face Completion for NIR-VIS Face Recognition.

Area of Science:

• Computer Vision
• Biometrics
• Artificial Intelligence

Background:

• Heterogeneous face recognition (HFR) matches images across different spectra, like Near-Infrared (NIR) and Visible (VIS).
• Existing HFR methods often synthesize VIS faces from NIR, but NIR data lacks visible lighting details, leading to incomplete images.
• This incompleteness poses challenges for accurate face matching due to self-occlusion and sensing gaps.

Purpose of the Study:

• To develop a high-resolution heterogeneous face synthesis method for improved NIR-VIS face recognition.
• To address the limitations of incomplete NIR data in face recognition tasks.
• To enhance the accuracy of matching NIR and VIS face images.

Main Methods:

• A novel approach models face synthesis as a combination of texture inpainting and pose correction.
• The texture inpainting component synthesizes VIS textures from NIR textures.
• The pose correction component maps NIR poses to frontal VIS poses, creating paired textures for an end-to-end deep network.

Main Results:

• The proposed network generates high-resolution VIS face images from NIR inputs.
• The method simplifies synthesis from one-to-many unpaired to one-to-one paired image translation.
• Experimental results demonstrate significant improvements in heterogeneous face recognition accuracy.

Conclusions:

• The integrated texture inpainting and pose correction effectively addresses spectral and pose discrepancies in HFR.
• The developed network achieves superior performance in both high-resolution face synthesis and recognition accuracy.
• This work offers a promising solution for robust face recognition across NIR and VIS domains.