Balancing Heterogeneous Image Quality for Improved Cross-Spectral Face Recognition.

Area of Science:

• Computer Vision
• Biometrics
• Image Processing

Background:

• Facial recognition across different modalities (visible light vs. infrared) and distances presents significant challenges.
• Infrared (IR) facial data often suffers from lower quality, particularly at longer ranges, hindering accurate matching with high-quality visible light images.

Purpose of the Study:

• To address the quality imbalance in heterogeneous facial imagery for improved cross-modality and cross-distance recognition.
• To propose and evaluate a novel cascaded face enhancement method to upgrade low-quality IR facial images.

Main Methods:

• A cascaded face enhancement approach combining the Block-Matching and 3D filtering (BM3D) denoising algorithm with a proposed deep-learning-based deblurring model, SVDFace.
• Experimental validation involving different IR bands (short-wave infrared - SWIR, near-infrared - NIR) and varying standoff distances.
• Comparison against a simple image downgrading approach and traditional/state-of-the-art deblurring methods.

Main Results:

• The proposed quality balancing approach consistently improved facial recognition performance across all tested IR bands and distances.
• The enhancement method was particularly effective for SWIR images captured at longer standoffs.
• The cascaded enhancement structure proved beneficial and necessary for optimal results.
• The SVDFace deblurring model demonstrated superiority over traditional and current deep learning deblurring techniques.

Conclusions:

• Upgrading low-quality IR facial images through cascaded enhancement is an effective strategy for improving cross-modality and cross-distance facial recognition.
• The proposed SVDFace model, inspired by singular value decomposition (SVD), offers an efficient, interpretable, and high-performing solution for deblurring challenging IR facial data.