Deep Cross-View Reconstruction GAN Based on Correlated Subspace for Multi-View Transformation

Area of Science:

• Computer Vision
• Biometrics
• Image Processing

Background:

• Visible spectrum (VIS) face identification is limited in poor lighting.
• Near-infrared (NIR) and thermal (TH) imaging offer alternatives but face domain shift challenges.
• Existing methods struggle with cross-domain face matching due to unique data distributions.

Purpose of the Study:

• To propose a novel image transformation framework for robust cross-domain face recognition.
• To enhance face matching accuracy between visible, near-infrared, and thermal imaging modalities.
• To generate high-quality, identity-preserving images across different spectral domains.

Main Methods:

• Feature extraction from input images.
• A transformation network generates target domain images with perceptual fidelity.
• A reconstruction network preserves original domain information.
• Framework applied to pix2pix and CycleGAN models (CRC-pix2pix, CRC-CycleGAN).
• Utilizes paired data considering feature correlation between domains.

Main Results:

• Generated high-quality images that preserve original face identity.
• Demonstrated superior performance in generated image face matching on TFW and BUAA NIR-VIS datasets.
• Achieved excellent evaluation metrics including SSIM, MSE, PSNR, and LPIPS.
• Introduced the CQUPT-VIS-TH dataset for thermal-visual face data.

Conclusions:

• The proposed framework effectively addresses cross-domain face recognition challenges.
• CRC-pix2pix and CRC-CycleGAN models show significant improvements in face matching.
• The approach is versatile and extensible to other image-to-image translation models.
• The new dataset facilitates further research in multi-modal face recognition.