cs.CVMar 2, 2026

Continuous Exposure-Time Modeling for Realistic Atmospheric Turbulence Synthesis

Authors: Junwei Zeng, Dong Liang, Sheng-Jun Huang, Kun Zhan, Songcan Chen

Organizations: Nanjing University of Aeronautics and Astronautics · School of Information Science and Engineering, Lanzhou University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Atmospheric turbulence significantly degrades long-range imaging by introducing geometric warping and exposure-time-dependent blur, which adversely affects both visual quality and the performance of high-level vision tasks. Existing methods for synthesizing turbulence effects often oversimplify the relationship between blur and exposure-time, typically assuming fixed or binary exposure settings. This leads to unrealistic synthetic data and limited generalization capability of trained models. To address this gap, we revisit the modulation transfer function (MTF) formulation and propose a novel Exposure-Time-dependent MTF (ET-MTF) that models blur as a continuous function of exposure-time. For blur synthesis, we derive a tilt-invariant point spread function (PSF) from the ET-MTF, which, when integrated with a spatially varying blur-width field, provides a comprehensive and physically accurate characterization of turbulence-induced blur. Building on this synthesis pipeline, we construct ET-Turb, a large-scale synthetic turbulence dataset that explicitly incorporates continuous exposure-time modeling across diverse optical and atmospheric conditions. The dataset comprises 5,083 videos (2,005,835 frames), partitioned into 3,988 training and 1,095 test videos. Extensive experiments demonstrate that models trained on ET-Turb produce more realistic restorations and achieve superior generalization on real-world turbulence data compared to those trained on other datasets. The dataset is publicly available at: github.com/Jun-Wei-Zeng/ET-Turb.

Figures & tables

Figure 1: Imaging results under atmospheric turbulence with different exposure-time. Short-exposure (e.g., 1 ms) primarily exhibits turbulence-induced tilt as turbulence state is effectively frozen. In contrast, long-exposure (e.g., 40 ms) integrates multiple turbulent states over time, resulting in significantly stronger blur .
Figure 2: Overview of the proposed exposure-time-dependent turbulence synthesis pipeline. The exposure-time τ\tau is treated as a continuous input parameter to both the ET-MTF and blur-width field formulations. By systematically varying τ\tau , our pipeline generates turbulence-degraded images with physically consistent blur characteristics from the input tilt image.
Figure 3: Comparison of MTF formulations at short and long-exposure limits. ET-MTF shows a smooth evolution of the MTF as exposure-time varies between the short and long-exposure limits.
Distance (m)Focal length (m)F-numberCn2C_{n}^{2} ( 10−14 m−2/310^{-14}\,\text{m}^{-2/3} )Height (m)Wind speed (m/s)Exposure-time (ms)
[30,100][0.1,0.3]{2.8,4}[50,300]{4,50}[1,3][0.5,8]
{2.8,4,5.6}[200,500]{100,200}[3,5][0.5,8]
[100,200][0.2,0.5]{2.8,4,5.6}[5,50]{200,400}[1,4][1,20]
{2.8,4,5.6,8}[20,100]{4,50}[2,6][0.5,10]
[200,400][0.3,0.5]{2.8,4,5.6,8}[2,30]{50,100}[2,5][1,20]
{4,5.6,8,11}[10,40]{10,50}[3,6][1,20]
Table 1: Parameter sampling ranges for ET-Turb dataset generation. Notation: [a,b][a,b] denotes uniform sampling from continuous interval; {v1,…,vn}\{v_{1},\ldots,v_{n}\} denotes uniform sampling from discrete set. All 12 configurations are sampled with equal probability.
Figure 4: Temporal evolution of turbulence degradation under Taylor’s frozen-flow hypothesis in video synthesis. We construct an extended turbulence degradation field along the wind direction vw\mathbf{v}_{w} . Temporal correlation between frames is achieved by displacing this field over time using the shift fvwtL\frac{f\mathbf{v}_{w}t}{L} .
Training DatasetTSR-WGAN [ 23 ]TMT [ 55 ]DATUM [ 53 ]MambaTM [ 54 ]
NIQE ↓\downarrowBRISQUE ↓\downarrowNIQE ↓\downarrowBRISQUE ↓\downarrowNIQE ↓\downarrowBRISQUE ↓\downarrowNIQE ↓\downarrowBRISQUE ↓\downarrow
TMT-dynamic [ 55 ]4.23152.5024.36158.5814.21954.9214.21755.062
ATSyn-dynamic [ 53 ]4.22454.4624.48359.7074.30859.1264.24756.876
ET-Turb4.19050.9814.22156.6914.20454.0704.21255.050
Table 2: Quantitative comparison of turbulence mitigation models trained on different synthetic datasets and evaluated on real turbulence data. Models trained on ET-Turb dataset achieve consistently lower NIQE and BRISQUE scores (lower is better), confirming superior realism and generalization capability compared to TMT-Dynamic [ 55 ] dataset and ATSyn-Dynamic [ 53 ] dataset.
Figure 5: Visual comparison of turbulence mitigation results on real data for MambaTM [ 54 ] models trained on different synthetic datasets. Models trained on ET-Turb dataset produce sharper and more natural restorations with fewer artifacts compared to those trained on TMT-Dynamic [ 55 ] dataset and ATSyn-Dynamic [ 53 ] dataset.
Figure 6: Visual ablation study on exposure modeling strategies for real turbulence restoration. The continuous ET-MTF formulation produces more natural textures and stable structures compared to fixed or binary exposure approaches.
Exposure StrategyNIQE ↓\downarrowBRISQUE ↓\downarrow
ET-Turb ( τ=1\tau=1 ms)4.35555.457
ET-Turb (MTF SE/LE{}_{\text{SE/LE}} )4.29755.123
ET-Turb (MTF ET{}_{\text{ET}} )4.21255.050
Table 3: Ablation study on exposure modeling strategies within the ET-Turb synthesis pipeline. Comparing fixed ( τ=1\tau=1 ms), binary (MTF SE/LE{}_{\text{SE/LE}} ), and continuous (MTF ET{}_{\text{ET}} ) exposure formulations demonstrates that continuous modeling yields the best perceptual quality on MambaTM [ 54 ] .
Figure 7: ET-Turb-Real dataset composition: multi-camera systems and diverse scene distribution. Real-world turbulence data is collected across three different imaging devices, providing comprehensive coverage of diverse environmental and anthropogenic scenes such as urban landscapes, entertainment venues, industrial infrastructure, traffic scenarios, and commercial establishments.
Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

MethodPSNR(dB) ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrow
TSR-WGAN [ 23 ]18.900.59270.5608
TMT [ 55 ]24.970.75070.3807
DATUM [ 53 ]22.850.64360.4700
MambaTM [ 54 ]23.770.66580.4490
Appendix
Table 4: Quantitative evaluation on the ET-Turb test set. TMT [ 55 ] achieves the best performance across all metrics, highlighting the effectiveness of its component-wise training strategy.
 Trained on ET-TurbTrained on ATSyn-Dynamic [ 53 ]
MethodPSNR(dB) ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrowPSNR(dB) ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrow
Noise-input20.57 (↓1.36%)0.58 (↓9.38%)0.57 (↑16.33%)20.57 (↓1.36%)0.58 (↓9.38%)0.57 (↑16.33%)
TMT [ 55 ]21.22 (↓0.38%)0.66 (↓1.49%)0.51 (↑2.21%)21.19 (↓0.42%)0.66 (↓2.94%)0.51 (↑2.37%)
DATUM [ 53 ]22.43 (↓0.80%)0.61 (↓4.69%)0.51 (↑2.89%)22.14 (↓1.07%)0.62 (↓7.46%)0.53 (↑8.16%)
MambaTM [ 54 ]22.12 (↓1.07%)0.64 (↓1.54%)0.52 (↑4.49%)22.10 (↓1.34%)0.63 (↓4.55%)0.52 (↑8.33%)
Appendix
Table 5: Robustness evaluation under exposure-dependent Gaussian noise added to the ET-Turb test set. Models trained on ET-Turb and ATSyn-Dynamic [ 53 ] are compared.
 r0r_{0} ±10%wind speed within 10-20m/s
MethodPSNR(dB) ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrowPSNR(dB) ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrow
Input20.71 (↓0.07%)0.78 (↓0.14%)0.34 (↑0.25%)20.43 (↓0.51%)0.76 (↓2.07%)0.36 (↑1.88%)
TMT [ 55 ]21.20 (↓0.06%)0.83 (↓0.14%)0.33 (↑0.12%)20.83 (↓0.82%)0.80 (↓1.19%)0.36 (↑2.22%)
DATUM [ 53 ]22.27 (↓0.11%)0.79 (↓0.17%)0.33 (↑0.18%)21.75 (↓0.87%)0.76 (↓1.48%)0.36 (↑2.37%)
MambaTM [ 54 ]21.57 (↓0.06%)0.78 (↓0.13%)0.36 (↑0.00%)21.19 (↓0.60%)0.75 (↓1.23%)0.38 (↑1.99%)
Appendix
Table 6: Sensitivity analyses with respect to r0r_{0} and wind speed. The Fried parameter r0r_{0} is perturbed by ±\pm 10%, while wind speed is increased from the training range (1–10 m/s) to 10–20 m/s.
Figure 8: Text recognition accuracy on the turbulence-text dataset [ 55 ] , evaluated using ASTER [ 44 ] .
Figure 9: Brightness distribution statistics of ET-Turb.
MethodsHardieChimittLauP2SOurs
Sec./frame119.635.883.130.353.24
Appendix
Table 7: Processing time for a 2562256^{2} frame.
Figure 11: Visualization of the turbulence synthesis pipeline across six parameter configurations. Each row corresponds to a distinct set of physical parameters from Tab. 11 . The columns show: (a) original undistorted image, (b) tilt with color-coded displacement field, (c) exposure-dependent blur, and (d) final synthesized turbulence combining both degradation components.
Figure 12: Generation of turbulent images under varying intensities and exposure-time using our data synthesis method. As the exposure-time and turbulence intensity increase, the degree of blur becomes more pronounced. The associated parameter set is distance = 500 m, focal length = 300 mm, F-number = 8, height = 50 m, and wind speed = 5 m/s.
Figure 13: Real-world turbulence mitigation results. Models trained on ET-Turb dataset demonstrate strong generalization to real atmospheric turbulence, successfully restoring fine details such as text and license plates while maintaining varying degrees of color fidelity.
Figure 14: Qualitative comparison of state-of-the-art turbulence mitigation methods trained on ET-Turb dataset. Each row corresponds to a specific method, showing results from two representative test scenes (left and right).

Explore similar work

CardsList
Feb 26, 2026

GSTurb: Gaussian Splatting for Atmospheric Turbulence Mitigation

Atmospheric turbulence causes significant image degradation due to pixel displacement (tilt) and blur, particularly in long-range imaging applications. In this paper, we propose a novel framework for atmospheric turbulence mitigation, GSTurb, which integrates optical flow-guided tilt correction and Gaussian splatting for modeling non-isoplanatic blur. The framework employs Gaussian parameters to represent tilt and blur, and optimizes them across multiple frames to enhance restoration. Experimental results on the ATSyn-static dataset demonstrate the effectiveness of our method, achieving a peak PSNR of 27.67 dB and SSIM of 0.8735. Compared to the state-of-the-art method, GSTurb improves PSNR by 1.3 dB (a 4.5% increase) and SSIM by 0.048 (a 5.8% increase). Additionally, on real datasets, including the TSRWGAN Real-World and CLEAR datasets, GSTurb outperforms existing methods, showing significant improvements in both qualitative and quantitative performance. These results highlight that combining optical flow-guided tilt correction with Gaussian splatting effectively enhances image restoration under both synthetic and real-world turbulence conditions. The code for this method will be available at https://github.com/DuhlLiamz/3DGS_turbulence/tree/main.

Hanliang Du, Zhangji Lu, Zewei Cai +3

Faculty of Science and Technology, University of Macau, Macau 999078, China · College of Physics and Optoelectronic Engineering, Shenzhen University, Shenzhen 518000, China · College of Physics and Optoelectronic Engineering, Shenzhen University, Shenzhen, 518060, China +4

Image RestorationImage DeblurringGaussian Splatting
AskPDF
May 25, 2026

D2^2Turb: Depth-Aware Simulation and Decoupled Learning for Single-Frame Atmospheric Turbulence Mitigation

Single-frame atmospheric turbulence mitigation is inherently ill-posed due to spatially varying blur coupled with non-rigid geometric distortion. Existing end-to-end approaches trained on flat-field simulations often struggle to balance texture recovery with geometric rectification. To overcome this limitation, we propose D2^2Turb, a unified framework that bridges physics-grounded simulation with explicitly decoupled restoration. First, we introduce a Depth-Aware Turbulence Synthesis protocol that incorporates scene depth into the phase-to-space formulation. This generates physically consistent, depth-dependent degradations and provides a crucial intermediate tilt supervision signal for disentangled learning. Building upon this simulation engine, D2^2Turb decomposes restoration into two interactive stages: texture deblurring and geometric rectification. The texture deblurring stage employs a deblurring backbone to recover fine-grained details while preserving geometric distortion for the subsequent rectification stage. To mitigate the information fragmentation commonly observed in cascaded designs, we further propose an Adaptive Structural Prior Injection (ASPI) mechanism that dynamically transfers deep structural representations from the deblurring module to guide dense flow prediction for spatial unwarping. Extensive experiments demonstrate that D2^2Turb achieves state-of-the-art performance on both synthetic and real-world datasets, with consistent improvements in both texture recovery and geometric fidelity. Our code and pre-trained models are publicly available at https://github.com/HertzDot222/D2Turb.

Zixiao Hu, Tianyu Li, Guoqing Wang +4

University of Electronic Science and Technology of China · Beijing Institute of Space Mechanics and Electricity · School of Physics, Northwest University Xi’an

Image Deblurring
AskPDF
May 20, 2026

ReMATF: Recurrent Motion-Adaptive Multi-scale Turbulence Mitigation for Dynamic Scenes

Atmospheric turbulence severely degrades video quality by introducing distortions such as geometric warping, blur, and temporal flickering, posing significant challenges to both visual clarity and temporal consistency. Current state-of-the-art methods are based on transformer, 3D architectures and require multi-frame input, but their large computational cost and memory usage limit real-time deployment, especially in resource-constrained scenarios. In this work, we propose ReMATF, a lightweight recurrent framework that restores videos using only two frames at a time while preserving spatial detail and temporal stability. ReMATF combines a multi-scale encoder-decoder with temporal warping and a motion-adaptive temporal fusion module that performs per-pixel fusion between the warped previous output and the current prediction to enhance coherence without enlarging the temporal window. This design reduces flicker, sharpens details, and remains efficient. Experiments on synthetic and real turbulence datasets show consistent improvements in PSNR/SSIM and perceptual quality (LPIPS), along with substantially faster inference than multi-frame transformer baselines, making ReMATF suitable turbulence mitigation in resource-constrained scenarios.

Zhiming Liu, Zhicheng Zou, Nantheera Anantrasirichai

Visual Information Laboratory, School of Computer Science, University of Bristol

Video Restoration
AskPDF