Boosting Restoration of Turbulence-Degraded Images With State Space Conditional Diffusion

Abstract

Recovering fine details from turbulence-distorted images is highly challenging due to the complex, spatially varying, and stochastic nature of the distortion process. Conventional multi-frame methods rely on extracting and averaging clear regions from pre-aligned frames, but their effectiveness is limited due to the rarity of “lucky regions”. In contrast, learning based methods have shown superior performance across various vision tasks. However, existing deep learning approaches still face key limitations: (1) they struggle to efficiently model the global context required for correcting pixel dispersion caused by spatially varying Point Spread Functions (PSFs); (2) they often overlook the physical formation of turbulence, particularly the spatial-frequency relationship between phase distortions and PSFs; and (3) they rely on deterministic architectures that fail to capture the inherent uncertainty in turbulence, leading to visually implausible outputs. To address these issues, we propose the Two-Stage Turbulence Removal Network (TSTRNet). The first stage uses a UNet-based generator built on the State Space Model to perform efficient, coarse global restoration. The second stage refines the output through a Denoising Diffusion Probabilistic Model, introducing stochasticity and edge-guided conditioning for detail enhancement and realism. Both stages incorporate frequency-domain processing to align with the physical characteristics of turbulence. Experimental results on multiple benchmark datasets demonstrate that TSTRNet achieves superior restoration performance compared to state-of-the-art methods, with strong generalization from synthetic to real-world scenarios.