Abstract
Single image deraining often struggles to generalize from synthetic training data to complex real-world environments. This paper introduces the High-order Collaborative Network (HCN), a three-tier architecture designed for efficient multi-scale feature sharing. To bridge the "synthetic-to-real" domain gap, we propose an online-update learning strategy that fine-tunes the model on real-world rainy images using self-supervision. Our approach achieves state-of-the-art performance on synthetic benchmarks while maintaining high efficiency and superior real-world generalization.
Introduction
Removing rain streaks from a single image is a classic ill-posed problem. While deep learning has advanced the field, two major hurdles remain: the underutilization of features across different image scales and the significant domain gap between synthetic training sets and real-world rain. Models trained on synthetic data often fail to handle the diverse physical properties and lighting effects of real rain. This work addresses these issues through a collaborative network structure and a dynamic adaptation mechanism.
Methodology
High-order Collaborative Networks (HCN)
The proposed HCN architecture consists of three sub-networks (Top, Middle, and Bottom) that work in tandem. While the bottom network extracts deep features, the top and middle networks provide "deep supervision" and transmit shallower features across scales. This collaborative approach ensures that rich information from intermediate layers is fully utilized, leading to better detail preservation than traditional single-stream networks.
Multi-Scale Constraints and Similarity Mining
To regularize the learning process, we employ Multi-Scale Compact Constraints (MSCC) using a Laplacian pyramid, forcing the model to learn useful features at multiple resolutions (e.g., 1/2 and 1/4 scales). Additionally, a Bidirectional Scale-Content Similarity Mining (BiSCSM) module identifies rain patterns by looking both coarse-to-fine and fine-to-coarse, capturing scale-invariant features more effectively.
Online-Update Learning
The "secret sauce" for real-world performance is the online-update strategy. When the model encounters real rainy images without ground truth, it uses its initial synthetic-trained prediction as a "pseudo-label." It then fine-tunes itself using a specialized loss function based on KL-divergence, ensuring background consistency while iteratively improving rain removal for that specific real-world domain.
Experimental Results
Performance and Efficiency
The model was evaluated on five synthetic benchmarks and a large collection of real-world images. On the challenging Rain200H dataset, our method achieved a PSNR of 29.985 dB, outperforming heavyweights like MPRNet (29.949 dB) and RCDNet (28.698 dB).
Crucially, the model is highly efficient. With only 2.04M parameters, it delivers superior results compared to much larger models like MSPFN (21.00M) and DCSFN (6.45M). This efficiency is visualized in the performance-parameter trade-off plot below.
Real-World Generalization
Qualitative results demonstrate that the online-update mechanism allows the model to effectively remove "haze-like" rain and preserve fine textures in real photos. This represents a significant improvement over models that rely solely on synthetic training, which often leave artifacts or fail to adapt to real-world rain densities.
Conclusion
By combining a high-order collaborative architecture with a self-supervised online-update strategy, we provide a robust solution for single image deraining. Our method balances high performance with computational efficiency, making it a strong candidate for practical applications. Future work will focus on optimizing the fine-tuning process for real-time use cases.
Summary updated