文献类型：期刊文献

英文题名：TDNet: degradation-aware comprehensive task decomposition for joint rain and haze removal

作者：Liang, Lei[1];Chen, Zhihua[1,2];Dai, Lei[1,2];Gao, Jiadan[1];Xia, Zhengran[1];Zhang, Yunyi[1]

机构：[1]East China Univ Sci & Technol, Dept Comp Sci & Engn, Shanghai 200237, Peoples R China;[2]East China Univ Sci & Technol, State Key Lab Ind Control Technol, Shanghai 200237, Peoples R China

年份：2026

卷号：42

期号：9

外文期刊名：VISUAL COMPUTER

收录：;EI(收录号：20262821071948);Scopus(收录号：2-s2.0-105043984095);WOS:【SCI-EXPANDED(收录号:WOS:001815148100002)】；

基金：This work was supported by the National Natural Science Foundation of China (Grant No. 62272164, Grant No. 62572188, and Grant No. 62306113) and the State Key Laboratory of Industrial Control Technology, China (Grant No: ICT2026A25).

语种：英文

外文关键词：Joint rain and haze removal; Composite weather degradation; Task decomposition; Physics-informed priors; Image restoration

摘要：Composite weather degradation, particularly the spatial coupling of rain and haze, degrades image quality and undermines the reliability of outdoor vision systems in both 2D and 3D visual tasks. Motivated by the distinct frequency characteristics of these weather artifacts, we propose a task decomposition network (TDNet), a physically grounded image restoration framework for joint rain and haze removal. Central to our method is the degradation-aware comprehensive task decomposition (DCTD) strategy, which reformulates the challenging restoration problem into three coordinated subtasks guided by physics-informed inductive biases. Specifically, we first devise an implicit neural deraining (IND) module that exploits the inherent spectral bias of implicit neural representations to suppress high-frequency rain artifacts. Subsequently, we introduce a prior-adaptive dehazing (PAD) module that models the atmospheric scattering process in the feature space to remove low-frequency haze effects. Finally, a scene restoration module (SRM) aggregates degradation-free features to recover high-fidelity image content. Extensive experiments on synthetic and real-world benchmarks show that TDNet compares favorably with 18 representative baselines. Codes are available at .