SD-PSFNet: Sequential and Dynamic Point Spread Function Network for Image Deraining

TL;DR

SD-PSFNet addresses image deraining under complex, multi-scale rain by introducing dynamic, learnable Point Spread Functions that adapt to local content. The method employs a three-stage sequential restoration framework with PSF-aware modules and a cross-stage gating mechanism to progressively refine rain removal while preserving details, guided by a physics-informed loss. Key contributions include a dynamic PSF mechanism, a suite of PSF-aware modules, and enhanced cross-stage fusion that deliver state-of-the-art PSNR/SSIM on multiple benchmarks and robust performance in dense rain conditions. The approach demonstrates that integrating physics-based degradation modeling with multi-stage learning yields practical, efficient deraining with strong generalization to real-world rain scenarios.

Abstract

Image deraining is crucial for vision applications but is challenged by the complex multi-scale physics of rain and its coupling with scenes. To address this challenge, a novel approach inspired by multi-stage image restoration is proposed, incorporating Point Spread Function (PSF) mechanisms to reveal the image degradation process while combining dynamic physical modeling with sequential feature fusion transfer, named SD-PSFNet. Specifically, SD-PSFNet employs a sequential restoration architecture with three cascaded stages, allowing multiple dynamic evaluations and refinements of the degradation process estimation. The network utilizes components with learned PSF mechanisms to dynamically simulate rain streak optics, enabling effective rain-background separation while progressively enhancing outputs through novel PSF components at each stage. Additionally, SD-PSFNet incorporates adaptive gated fusion for optimal cross-stage feature integration, enabling sequential refinement from coarse rain removal to fine detail restoration. Our model achieves state-of-the-art PSNR/SSIM metrics on Rain100H (33.12dB/0.9371), RealRain-1k-L (42.28dB/0.9872), and RealRain-1k-H (41.08dB/0.9838). In summary, SD-PSFNet demonstrates excellent capability in complex scenes and dense rainfall conditions, providing a new physics-aware approach to image deraining.

Figures (7)

• Figure 1: Overview of the SD-PSFNet. The left panel shows the serialized overall restoration framework of SD-PSFNet, including three stages: input Stage In, multiple Stage Mid, and final restoration output ORStage. The right panels detail our key components: PSF Channel Reducer, PSF-Aware Attention, CA/PSF Block, and Multi-Scale PSF Head Module. The network integrates multi-scale dynamic PSF mechanisms for modeling, effectively handling multi-scale rain removal. Input images undergo serialized restoration through three stages with specialized PSF-aware processing.
• Figure 2: Qualitative deraining performance comparisons on Rain100L yang2019jointand Rain100H yang2019jointdatasets. Our SD-PSFNet achieves competitive visual results comparable to the other SOTA method.
• Figure 3: Comparison of restoration results across different stages on three datasets (from top to bottom): Rain100H yang2019joint and two RealRain-1k-H li2022toward samples. Stage 1 represents Stage In, Stages 2-4 correspond to Stage Mid, and Stage 5 shows the final restoration from ORStage.
• Figure 4: Example images used in additional visualization experiments.
• Figure 5: PSNR drop when disabling pathways of gate-based feature fusion between different stages and components.