Introduction
Single-image de-raining is a challenging "ill-posed" problem where an algorithm must distinguish between actual scene details and overlapping rain streaks. Most modern solutions rely on Deep Convolutional Neural Networks (CNNs). However, standard CNNs are inherently local operators, processing small neighborhoods of pixels at a time.
Rain streaks are often spatially long and complex, requiring a global perspective to identify patterns across the entire image. While "non-local" modules can capture these long-range dependencies, they are typically memory-intensive and slow. This paper introduces a Dual Graph Convolutional Network (GCN) designed to capture global context—both spatially and across feature channels—without the massive computational overhead of traditional non-local methods.
Methodology
The proposed architecture utilizes a symmetrically skip-connected structure containing a "Basic Unit" repeated 11 times. This unit is designed to synergize local and global information through three primary components:
- Spatial GCN (sGCN): This module treats pixels as nodes in a graph, allowing every pixel to communicate with every other pixel regardless of distance. By employing a matrix re-ordering trick based on the associative rule, the authors reduce computational complexity from $O((HW)^2)$ to $O(HW)$, making it significantly faster than standard non-local modules.
- Dilated Convolutional Module (DCM): To handle rain streaks of varying sizes, the DCM uses parallel paths of dilated convolutions. This enables the network to capture local features across multiple receptive fields (ranging from 3x3 to 13x13) simultaneously.
- Channel GCN (cGCN): While the spatial GCN focuses on positional relationships, the channel GCN models interdependencies between different feature channels. This helps the network suppress rain-related features while preserving essential scene details.
Experimental Evaluation
The model was evaluated on several widely recognized benchmarks, including synthetic datasets (Rain12, Rain200L/H, DDN-Data, and DID-Data) and the large-scale real-world SPA-Data set, which contains approximately 640,000 image pairs. The network was trained end-to-end using a Mean Absolute Error (MAE) loss function to balance rain removal with detail preservation.
Key Results and Insights
Quantitative and Qualitative Performance
The Dual GCN model consistently outperformed 14 baseline methods (including NLEDN, RESCAN, and PReNet) across all benchmark datasets. On the challenging Rain200H dataset, it achieved a PSNR of 31.15 dB. Visually, the model excels at preserving fine textures—such as fabric weaves or distant architectural edges—that competing models often blur during the de-raining process.
Efficiency and Orthogonal Modeling
A key insight of this work is the use of orthogonal global modeling. By addressing spatial and channel relations separately but within the same framework, the network gains a richer representation of the image. Furthermore, the authors demonstrate that GCNs, when implemented efficiently, provide a superior alternative to self-attention-style modules for non-local operations in computer vision.
Conclusion
The Dual GCN framework provides a robust and efficient solution for single-image de-raining by effectively combining local multi-scale convolutions with global graph reasoning. While currently focused on static images, the authors suggest that this framework could be extended to video de-raining or adapted for other low-level vision tasks like denoising and super-resolution. Future work may also address extreme weather conditions, such as rain accumulation and mist, which require additional dehazing components.
Summary updated