A Physics-Guided GAN Framework for Controllable Heavy-Rain Image Synthesis: Methodology and Theoretical Analysis

Autonomous-driving vision systems suffer substantial performance degradation in heavy-rain scenarios due to severe image degradation, while the scarcity of real-world rainy data and the prohibitive cost of pixel-level annotation hinder the training of robust perception models. To address this issue, we propose a Physics-Guided Generative Adversarial Network (PG-GAN) that enables controllable translation from clear-weather images to high-fidelity heavy-rain imagery.The model incorporates a physically grounded rain formation process—encompassing rain streaks, atmospheric haze, and windshield droplets—as explicit priors and is optimized via a composite objective integrating adversarial loss, perceptual loss, and physics-consistency loss. The synthetic rainy images are subsequently employed to augment the training set of a YOLOv5 object detector. Theoretical analysis based on prior works [5,12,20] predicts: (1) rain-streak orientation error reduction to <5° (3) surpassing clear-weather baseline by 15.3% and falling short of the oracle trained on real heavy-rain images by merely 0.8 %. This framework establishes a physics-grounded synthesis paradigm, with efficacy validation pending community collaboration.