MODEM: A Morton-Order Degradation Estimation Mechanism for Adverse Weather Image Recovery

TL;DR

MODEM tackles adverse-weather image restoration by explicitly estimating spatially varying degradation using a Dual Degradation Estimation Module (Z0, Z1) and guiding restoration with a Morton-Order 2D-Selective-Scan Module (MOS2D). The two-stage training pipeline leverages global and local degradation priors to adaptively modulate feature processing, yielding context-aware restoration through Degradation-Adaptive Feature Modulation and Degradation-Selective Attention Modulation. Empirical results across Snow100K, Outdoor-Rain, RainDrop, and real-world snow datasets show state-of-the-art performance on average, with notable gains in PSNR, SSIM, and perceptual quality metrics like LPIPS, Q-Align, and MUSIQ. The approach demonstrates strong generalization to real-world degraded images and emphasizes the value of explicit degradation estimation for robust, adaptable restoration.

Abstract

Restoring images degraded by adverse weather remains a significant challenge due to the highly non-uniform and spatially heterogeneous nature of weather-induced artifacts, e.g., fine-grained rain streaks versus widespread haze. Accurately estimating the underlying degradation can intuitively provide restoration models with more targeted and effective guidance, enabling adaptive processing strategies. To this end, we propose a Morton-Order Degradation Estimation Mechanism (MODEM) for adverse weather image restoration. Central to MODEM is the Morton-Order 2D-Selective-Scan Module (MOS2D), which integrates Morton-coded spatial ordering with selective state-space models to capture long-range dependencies while preserving local structural coherence. Complementing MOS2D, we introduce a Dual Degradation Estimation Module (DDEM) that disentangles and estimates both global and local degradation priors. These priors dynamically condition the MOS2D modules, facilitating adaptive and context-aware restoration. Extensive experiments and ablation studies demonstrate that MODEM achieves state-of-the-art results across multiple benchmarks and weather types, highlighting its effectiveness in modeling complex degradation dynamics. Our code will be released at https://github.com/hainuo-wang/MODEM.git.

Figures (15)

• Figure 1: Comparison of various image scanning methods. (a) Raster Scanning. (b) Continuous Scanning. (c) Local Scanning. (d) Morton Scanning, which can better preserve spatial locality of neighboring pixels in the resulting 1D sequence, beneficial for capturing contextual information.
• Figure 2: Connection between MODEM and SSM.
• Figure 3: With respect to a sample (a), (b)-(d) visualize the long-range $CAh_{k-1}$, (c) local $CBx_k$, and (d) output of MOS2D $y_k$, respectively. More cases can be found in Appendix \ref{['sec:more_fea']}
• Figure 4: (a) Overall architecture of MODEM. (b) The DDEM for extracting global descriptor $Z_0$ and adaptive degradation kernel $Z_1$ degradation priors. (c) The MDSL incorporating the core MOS2D module (d) within a residual block. The blue-colored components indicate elements exclusive to the first training stage. $N$, $M_1$, $M_2$ denote the number of the corresponding module, respectively.
• Figure 4: Deraining & Dehazing Task