Gaussian Process-Driven Semi-Supervised Single-Image Rain Removal: Enhancing Real-Scene Generalizability

This paper proposes a semi-supervised single-image rain removal method using Gaussian processes to decouple rain components and background features. Existing methods often fail to generalize to real scenes due to synthetic data's limited diversity in rain direction and density. To address this, we integrate synthetic and real rainy images, where Gaussian processes model synthetic intermediate features to generate pseudo-labels for real image supervision. A two-stage encoder–decoder architecture with squeeze-and-excitation residual and context feature fusion modules enhances feature disentanglement. Experimental results on both synthetic and real datasets demonstrate superior performance, achieving a peak signal-to-noise ratio of 26.11 dB and structural similarity of 0.89 on synthetic images, while preserving more background details and effectively supporting downstream tasks like object segmentation.