Abstract
Single Image Deraining (SID) is a critical task for enhancing visibility in outdoor vision systems. Traditional deep learning approaches often require paired "rainy/clean" images for training, which are difficult to obtain in real-world settings. This paper introduces CCLGAN, a novel unsupervised framework that utilizes Cycle Contrastive Learning (CCL) and Location Contrastive Learning (LCL) to achieve high-quality deraining without paired data. By integrating a pre-trained CLIP model as a semantic anchor and employing a dual-branch contrastive strategy, CCLGAN preserves high-level semantic meaning and spatial consistency. Evaluations on synthetic and real-world benchmarks show that CCLGAN significantly outperforms state-of-the-art unsupervised methods and rivals supervised models in visual quality and quantitative metrics.
Introduction
Single Image Deraining (SID) aims to recover a clean background from an image obscured by rain. While supervised models have shown success, their reliance on paired datasets limits their applicability to real-world scenarios where such pairs are unavailable. Unsupervised methods like CycleGAN have been proposed but often suffer from "channel pollution," where background details are lost or rain information is incorrectly preserved. CCLGAN addresses these challenges by shifting from pixel-level consistency to feature-based contrastive learning, ensuring the model understands both the content and the spatial structure of the scene.
Methodology
The CCLGAN architecture is built upon two primary components designed to decouple the rain layer from the semantic background:
- Cycle Contrastive Learning (CCL): This component ensures semantic consistency across the deraining process. It is divided into two branches: Intra-CCL, which utilizes the semantic latent space of a pre-trained CLIP model to pull the reconstructed image toward clean semantic anchors, and Inter-CCL, which operates in a discriminant latent space to isolate and filter the rain layer.
- Location Contrastive Learning (LCL): To maintain spatial integrity, LCL maximizes mutual information between corresponding patches in the input and output images. This prevents the "hallucination" of objects and ensures that background elements remain in their correct coordinates.
Experimental Evaluation
The model was evaluated on the RainCityscapes (synthetic) and SPA (real-world) datasets. Quantitatively, CCLGAN achieved a PSNR of 29.17 dB on RainCityscapes, a significant improvement over the previous unsupervised state-of-the-art, ANLCL (27.42 dB). On the SPA dataset, CCLGAN's performance was found to be comparable to older supervised models, demonstrating its robustness in real-world conditions.
Qualitative results illustrate that CCLGAN produces sharper images with fewer artifacts compared to other unsupervised methods like CUT and DerainCycleGAN. The model effectively removes both heavy rain streaks and the "veiling" haze often associated with precipitation, as shown in the comparison below.
Conclusion
CCLGAN represents a significant step forward in unsupervised image deraining. By leveraging CLIP as a semantic teacher and employing localized contrastive losses, the framework overcomes the limitations of traditional pixel-based unsupervised learning. While the model introduces some computational overhead due to the CLIP integration, its ability to generalize to real-world data without paired training sets makes it a powerful tool for outdoor vision applications. Future research will explore extending this universal unsupervised strategy to other adverse weather conditions such as fog or snow.
Summary updated