Multi-input multioutput dual estimation network for atmospheric turbulent image restoration
Lizhen Duan, Meihui Li, Dongxu Liu, Yunfeng Liu, Jianlin Zhang · 2025
In this paper, we present a deep learning-based approach to restore single-frame aircraft images degraded by atmospheric turbulence. First, we simulate hundreds of different turbulent blur kernels with different turbulence levels using Von Kármán power spectrum model. Then, we construct training image pairs by pairing clear images sourced from the FGCV-Aircraft dataset with their corresponding blurry counterparts, which had been convolved with the synthetic blur kernels. The synthetic dataset is utilized to train a multi-input multi-output dual estimation model to learn the clear image and the turbulent blur kernel. The degraded images serve as network inputs, while the corresponding clear images and blur kernels serve as network targets. The pixel-wise content loss and frequency reconstruction loss are used as the estimation losses for both clear image and blur kernel image. The experimental results on test and real-world turbulent images demonstrate the efficacy of the proposed method.