- Recent days: Results of compared methods; Pretrained weights and inference/evaluation codes.
- Jan 23, 2026: Initialization of the repository.
Simplified abstract: This paper has proposed a task decomposition network for joint rain and haze removal in single images, featuring a degradation-aware comprehensive task decomposition (DCTD) based on the distinct properties of weather artifacts. Three subtasks are performed: (1) High-frequency rain removal based on implicit neural deraining module (IND); (2) Low-frequency haze removal through prior adaptive dehazing module (PAD); (3) Weather-degraded background reconstruction via scene restoration module (SRM). Extensive experiments are conducted to demonstrate the efficiency of its designs. Seven benchmarks are included for comparisons or ablations, including RainCityscapes, RainCityscapes-pp, RainhazeSynscapes, Rain200H, RW2AH, SemiSIRR, and REAL-RAIN. PSNR, SSIM, and LPIPS scores are measured for reference-based IQA. For reference-free IQA, we employ NIQE and CLIP-IQA.
More details can be found in the TXT or YAML files from the envs folder.
We will first sort and upload predicted results of compared methods and ours on involved benchmarks. Afterwards, the pretrained weights of our method will be uploaded. Please stay tuned.
Our work is built upon the codebase of SFNet, TransMamba, and C2PNet, and we sincerely thank them for their contributions.