Adaptive diffusion models for overcoming data scarcity in long-distance face recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Long-distance face recognition (FR) is challenging due to image degradation and limited training data in surveillance.
• Existing methods struggle with low-resolution and poor-quality facial images captured from afar.

Purpose of the Study:

• To introduce Face-Aware Diffusion (FADiff), an Adaptive Diffusion Model (ADM) designed to overcome data scarcity and improve long-distance FR.
• To enhance facial image quality and maintain identity consistency for surveillance and security applications.

Main Methods:

• FADiff integrates a Face Condition Embedding Module (FCEM) using ArcFace-trained ResNet101 and MLP-Mixer for identity conditioning.
• A Face-Aware Initial Estimator (FAIE) with modified SwinIR and hierarchical attention provides structural initialization.
• An ADM with Feature-wise Linear Modulation (FiLM) reconstructs high-fidelity, identity-consistent faces.

Main Results:

• FADiff achieved significant improvements on the WIDER-FACE dataset: 27.84 dB PSNR, 0.821 SSIM, 0.743 ArcFace similarity, and 0.612 detection AP@0.5.
• Outperformed the best baseline (DiffBIR) by 6.3% to 11.9% on the Hard subset, with statistically significant results (p < 0.001).
• Demonstrated excellent scalability and efficiency, with practical training and inference times suitable for real-world deployment.

Conclusions:

• FADiff effectively addresses data scarcity and image degradation in long-distance FR.
• The model offers superior performance and scalability for surveillance and security applications.
• FADiff represents a practical advancement in robust facial recognition technology.