TY - JOUR
T1 - Receptive field enhanced Transformer with multi-attention mechanisms for single-image deraining
AU - Li, Jingyu
AU - Zhang, Zhenhai
AU - Huang, Yilei
AU - Wang, Ruosong
AU - Chen, Siyu
N1 - Publisher Copyright:
© 2026 Elsevier B.V.
PY - 2026/11/4
Y1 - 2026/11/4
N2 - Single-image deraining under mixed degradation caused by rain streaks and raindrops remains a significant challenge in real-world outdoor image restoration, where clear visual content is important for preserving scene details and supporting subsequent image analysis. This difficulty stems from the different spatial characteristics and degradation patterns of rain streaks and raindrops. To overcome these issues, we propose a Transformer-based deraining network that captures global context while preserving fine-grained details. The architecture adopts a symmetrical encoder–decoder design with skip connections and integrates two key modules. Specifically, the first module, Receptive Field Enhanced Window-based Attention, is designed to address mixed rain degradation by enhancing standard window-based attention with adaptive window scoring and cross-window interactions. This design enables the modeling of long-range spatial dependencies between rain streaks and raindrops, which is essential for distinguishing complex rain artifacts from background textures, while avoiding the high computational costs associated with vanilla global attention. The second module, termed Dual-Stream Deformable Self-Attention, performs sparse and structure-aware sampling around degraded regions. It employs separate attention branches for rain streaks and raindrops, improving reconstruction accuracy. Together, these modules enable the network to effectively identify and restore various types of rain-induced degradations with high fidelity. Experimental results demonstrate that the proposed method exhibits favorable overall effectiveness compared to existing rain removal methods.
AB - Single-image deraining under mixed degradation caused by rain streaks and raindrops remains a significant challenge in real-world outdoor image restoration, where clear visual content is important for preserving scene details and supporting subsequent image analysis. This difficulty stems from the different spatial characteristics and degradation patterns of rain streaks and raindrops. To overcome these issues, we propose a Transformer-based deraining network that captures global context while preserving fine-grained details. The architecture adopts a symmetrical encoder–decoder design with skip connections and integrates two key modules. Specifically, the first module, Receptive Field Enhanced Window-based Attention, is designed to address mixed rain degradation by enhancing standard window-based attention with adaptive window scoring and cross-window interactions. This design enables the modeling of long-range spatial dependencies between rain streaks and raindrops, which is essential for distinguishing complex rain artifacts from background textures, while avoiding the high computational costs associated with vanilla global attention. The second module, termed Dual-Stream Deformable Self-Attention, performs sparse and structure-aware sampling around degraded regions. It employs separate attention branches for rain streaks and raindrops, improving reconstruction accuracy. Together, these modules enable the network to effectively identify and restore various types of rain-induced degradations with high fidelity. Experimental results demonstrate that the proposed method exhibits favorable overall effectiveness compared to existing rain removal methods.
KW - Deep learning
KW - Deformable self-attention
KW - Image deraining
KW - Transformer
KW - Window-based attention
UR - https://www.scopus.com/pages/publications/105052812821
U2 - 10.1016/j.knosys.2026.117168
DO - 10.1016/j.knosys.2026.117168
M3 - Article
AN - SCOPUS:105052812821
SN - 0950-7051
VL - 353
JO - Knowledge-Based Systems
JF - Knowledge-Based Systems
M1 - 117168
ER -