Physics-Guided Generative Adversarial Networks for Robust Low-Light Image Enhancement

Low-light image enhancement is a fundamental preprocessing step for many applications in computer vision, such as surveillance, autonomous navigation, and medical imaging. Most state-of-the-art deep learning–based enhancement approaches, among which Generative Adversarial Networks hold the majority, have impressive visual quality but generally lack robustness and generalisation because of their fully data-driven nature. This paper proposes the Physics-Guided Generative Adversarial Network, which incorporates physical image formation principles, among which are illumination degradation, sensor noise, and camera response, into the GAN learning framework. By embedding a physics-based constraint into both generator architecture and loss function, our proposed method yields enhanced perceptual quality, noise suppression, and structural fidelity across diverse low-light conditions. Experimental results on benchmark datasets demonstrate that PG-GAN outperforms state-of-the-art enhancement techniques in large margins along PSNR, SSIM, and perceptual metrics while exhibiting superior robustness across unseen illumination scenarios.