Long-range Turbulence Mitigation: A Large-scale Dataset and A Coarse-to-fine Framework

TL;DR

This work tackles the challenge of long-range atmospheric turbulence, which causes severe distortions in far-field imaging. It introduces the RLR-AT dataset with 1500 real sequences spanning 1–13 km and proposes a coarse-to-fine framework (CDSP) that combines a frequency-aware reference frame with a subspace-based low-rank tensor refinement to jointly leverage dynamic turbulence priors and static background priors. The approach comprises a frequency-aware reference frame to improve registration and a low-rank tensor refinement to preserve details while correcting residual misalignments, followed by a data-driven deblurring step. Experiments on synthetic and real benchmarks show that CDSP outperforms state-of-the-art methods, with ablation confirming the complementary roles of FRF and SLRTR and their robustness to long-range distortions, highlighting substantial potential for long-range imaging applications.

Abstract

Long-range imaging inevitably suffers from atmospheric turbulence with severe geometric distortions due to random refraction of light. The further the distance, the more severe the disturbance. Despite existing research has achieved great progress in tackling short-range turbulence, there is less attention paid to long-range turbulence with significant distortions. To address this dilemma and advance the field, we construct a large-scale real long-range atmospheric turbulence dataset (RLR-AT), including 1500 turbulence sequences spanning distances from 1 Km to 13 Km. The advantages of RLR-AT compared to existing ones: turbulence with longer-distances and higher-diversity, scenes with greater-variety and larger-scale. Moreover, most existing work adopts either registration-based or decomposition-based methods to address distortions through one-step mitigation. However, they fail to effectively handle long-range turbulence due to its significant pixel displacements. In this work, we propose a coarse-to-fine framework to handle severe distortions, which cooperates dynamic turbulence and static background priors (CDSP). On the one hand, we discover the pixel motion statistical prior of turbulence, and propose a frequency-aware reference frame for better large-scale distortion registration, greatly reducing the burden of refinement. On the other hand, we take advantage of the static prior of background, and propose a subspace-based low-rank tensor refinement model to eliminate the misalignments inevitably left by registration while well preserving details. The dynamic and static priors complement to each other, facilitating us to progressively mitigate long-range turbulence with severe distortions. Extensive experiments demonstrate that the proposed method outperforms SOTA methods on different datasets.

Figures (10)

• Figure 1: Visual examples of turbulence mitigation on proposed real-world long-range atmospheric turbulence benchmark RLR-AT. The proposed method could effectively handle the long-range turbulence with severe distortions.
• Figure 2: Illustration of the proposed dataset RLR-AT. (a) Long-range imaging with larger focal length lens through turbulence. (b) Typical turbulence with diverse long-distance conditions. (c) Statistics of distance and scene of the proposed benchmark.
• Figure 3: Comparison between Temp Avg and proposed FRF. (a) Comparison of registration performance between Temp Avg and FRF. Temp Avg fails to achieve precise registration since it suffers from blur, while FRF achieves more accurate registration. (b) Comparison between the construction of Temp Avg and FRF. Temp Avg assigns equal weight to all intensities at a certain position. However, the output is dissimilar with original GT, since the less frequently occurring intensities make a negative contribution. In contrast, we argue that the higher the frequency, the greater the weight, as the most frequently occurring pixel at the position is closer to the original GT.
• Figure 4: Analysis of pixel motion statistical priors using checkerboard corners.(a) visualizes the clean and distorted checkerboard and the motion field of corners. (b) shows the motion trajectory of corners along the temporal axis, with an extended display of two randomly selected corners across 200 frames. (c) performs a statistic of corner motions and the motions of corners approximately conform to a zero-mean gaussian distribution, indicating that corners most frequently occur in their original positions.
• Figure 5: Low-rank property analysis of different sequences. (a) Visualization of selected section lines. (b) Singular value curves of corresponding sequences.