Substantial Image Deraining with Realistic Generalization Through Gaussian Process-Driven Semi-Supervised Analysis

Image deraining has seen significant advancements with CNN-based methods achieving impressive results in reconstruction accuracy and visual quality. However, these methods are often constrained by their dependence on fully labelled training data, limiting their generalization capabilities for real-world scenarios. To address this gap, we propose a novel Gaussian Process-Driven Semi-Supervised Learning Framework (GPSSL) for robust image deraining, titled Substantial Image Deraining with Realistic Generalization Through Gaussian Process-Driven Semi-Supervised Analysis. In order to provide pseudo-ground-truth for improved generalization, this approach models the latent space of unlabeled real-world images using Gaussian Processes and uses synthetic datasets for supervised training. Comprehensive tests on benchmark datasets like Rain800, Rain200L, and DDN-SIRR show how effective our method is. For example, our approach outperforms state-of-the-art techniques like JORDER (PSNR: 30.1 dB, SSIM: 0.87) and RESCAN (PSNR: 31.2 dB, SSIM: 0.89) on the Rain800 dataset, achieving a PSNR of 32.4 dB and SSIM of 0.91. Similarly, on the Rain200L dataset, GPSSL achieves a PSNR of 36.8 dB, surpassing competitive baselines like DID-MDN (PSNR: 34.6 dB). The results highlight a 15-20% improvement in generalization performance when incorporating unlabelled real-world data. Our proposed framework not only bridges the gap between synthetic and real-world scenarios but also sets a new benchmark in leveraging unlabelled data for image deraining tasks, paving the way for future advancements in semi-supervised learning techniques for adverse weather restoration applications.