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

Jiayu Wang, Haoyu Bian, Haoran Sun, Shaoning Zeng · 2025

Read SD-PSFNet: Sequential and Dynamic Point Spread Function Network for Image Deraining by Jiayu Wang, Haoyu Bian, Haoran Sun, Shaoning Zeng (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-17993.txt.

Introduction

Jiayu Wang, Haoyu Bian, Haoran Sun, and Shaoning Zeng present SD-PSFNet, a network for removing rain from images. Rain removal is hard because streaks vary in scale and tangle with the scene behind them. Borrowing from optics, the design uses point spread function components, which model how a point of light smears, to expose the degradation process that rain imposes on a photo.

The network restores the image in three linked stages, re-estimating the rain effect at each stage and refining the result step by step. A gated fusion step passes the best features between stages. The authors present the system as a way to separate rain from background more cleanly by modelling the physics of the streaks rather than treating removal as pure pattern matching.

Catalog record

CREATORJiayu Wang, Haoyu Bian, Haoran Sun, Shaoning Zeng
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-17993

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

SD-PSFNet: Sequential and Dynamic Point Spread Function Network for Image Deraining Jiayu Wang, Haoyu Bian, Haoran Sun, Shaoning Zeng 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. Code — https://github.com/Aster-1024/SD-PSFNet Introduction Image deraining is a fundamental restoration task aiming to recover rain-free images from rainy ones. This task is crucial for downstream applications ( 25 ) such as object detection ( 29 ) , autonomous driving ( 26 ) , and surveillance systems ( 9 ) , as rain-degraded images suffer from reduced visibility, decreased contrast, and detail loss that significantly impact visual processing performance. In recent years, deep learning-based deraining methods have made significant progress ( 28 ) . From early CNN-based models to approaches incorporating Transformer architectures such as Restormer ( 41 ) and Uformer ( 32 ) , as well as GAN-based techniques like SSCGAN ( 24 ) and LFI-based ( 10 ) , these approaches have demonstrated remarkable capabilities in image deraining tasks. The primary challenges in image deraining arise from the complexity, multi-scale nature, and high coupling of raindrop distributions with background scenes. Most existing methods primarily focus on learning mapping relationships from data, overlooking the physical and optical properties of raindrop formation ( 38 ) . This neglect leads to an inability to dynamically adjust to different rain patterns, resulting in insufficient deraining and lack of explainability. Although Transformer-based architectures, GAN-based models, and diffusion models demonstrate superior performance in image restoration tasks, their massive parameter counts and computational overhead hinder deployment in real-time applications ( 6 ) . Additionally, existing methods attempt to incorporate domain knowledge through physics-aware approaches, such as leveraging rain streaks’ frequency domain characteristics as inductive bias. However, traditional explicit physical modeling relies on fixed prior assumptions, such as static PSF templates that struggle to capture the diversity of rain degradation, leading to overlooking the multi-scale distribution of rain streaks ( 45 ) and optical differences of raindrops. To address these issues, we refocus our attention on CNN architecture and propose a sequential modeling-based deraining network featuring a novel physics-aware mechanism that relies on multi-scale, data-driven dynamic Point Spread Function (PSF) ( 23 ) prediction. Based on this mechanism, we propose SD-PSFNet, constructing multiple innovative PSF-related modules integrated into a powerful multi-stage architecture, inspired by successful progressive restoration paradigms like MPRNet ( 40 ) . Additionally, we implement an adaptive gating fusion mechanism to achieve sequential modeling, facilitating the transmission and optimization of cross-stage information within the multi-stage architecture. Experiments show SD-PSFNet achieves state-of-the-art deraining performance across public datasets while maintaining reasonable computational efficiency. Compared to the baseline MPRNet architecture, our novel physics-aware design delivers a significant 5.04dB PSNR improvement (13.5% gain) on RealRain-1k-L, with particular effectiveness in complex scenes and dense rain conditions. The key contributions are: • A Novel PSF-Integrated Deraining Approach : Multiple integrated PSF mechanism modules dynamically simulate raindrops’ optical effects, enhancing representation of complex rain patterns. • Sequential Model Design : Multi-level gated fusion strategies enable full transmission of physical prior information across multiple stages, repeatedly and dynamically refining the estimation of image degradation. • Extensive Experimental Evaluation : Multiple benchmark evaluations, including comparative tests and ablation studies, confirm our method’s superior performance. Related Work Image Deraining Image deraining aims to recover rain-free images from rainy ones, representing a specific instance of reversing image degradation. Traditional methods enhanced interpretability by modeling rain’s optical properties through image decomposition ( 17 ) and frequency domain analysis ( 44 ) . However, these approaches struggle with complex scenes, causing over-smoothing or residual rain streaks. Deep learning has emerged as the primary approach for image deraining, evolving from CNN-based local feature extraction to global modeling and generative techniques. Early CNNs like DANet ( 15 ) and DerainRLNet ( 1 ) processed rain hierarchically but suffered from limited receptive fields. Later, multi-scale networks such as RESCAN ( 19 ) and DID-MDN ( 42 ) along with physically-guided networks like KGCNN ( 31 ) enhanced adaptability through multi-resolution feature fusion and motion blur modeling. Restormer ( 41 ) applied Transformer ( 27 ) to low-level vision, achieving linear-complexity long-range dependency modeling via cross-channel self-attention. Generative approaches including Attentive GAN ( 21 ) produced realistic rain-free images through adversarial training and attention mechanisms, though facing mode collapse issues. Semi-supervised and unsupervised methods such as UD-GAN ( 16 ) and DerainCycleGAN ( 33 ) addressed the scarcity of real rain annotations by leveraging unpaired data, while DCDGAN ( 8 ) utilized diffusion models for enhanced realism. However, these methods still lack multi-scale physics-aware modeling and struggle to balance computational efficiency with deraining performance. Physics-Aware Image Restoration Physics-aware approaches are widely used in image restoration, integrating optical imaging’s physical mechanisms into models ( 35 ) . Fusion strategies combining explicit physical modeling with data-driven deep learning approaches have become mainstream methods. For example, ReDNet ( 34 ) constructed an optically-guided Transformer network that precisely separates stains from background information. In applications focusing on rainy scenes, the ReDT-Det ( 2 ) network employs Illumination Fusion Differential Transformer Blocks (IFDTB) to suppress noise and restore details in dark regions, significantly enhancing vehicle detection capabilities in nighttime rainy and foggy conditions. Similarly, researchers have proposed a Dual-Scale Prior (DSP) ( 39 ) model that combines physical priors such as non-local self-similarity with the powerful representational capabilities of pre-trained deep denoising networks, achieving robust removal of degradation factors like rain streaks. However, multi-scale rain streak removal still requires further exploration and enhancement. 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. Method This section details SD-PSFNet’s implementation, including theoretical validation of PSF and the innovative module and framework design. Dynamic PSF Mechanism Image degradation processes can typically be modeled as linear space-invariant systems using Point Spread Function (PSF), but real rainy image degradation is more complex, with traditional models struggling to accurately characterize spatially varying degradation fields K(x,y) . We transform

Read the complete text