Multi-Scale Atmospheric Turbulence Blind Deblurring
Yang Liu, Zhisheng Gao · 2024
When observing space targets, atmospheric turbulence causes geometric distortion and spatiotemporal changes in the ground-to-air imaging process due to the random motion of the turbulent medium and the superposition of multiple external disturbance factors. Existing methods either perform deblurring and denoising completely independently, or ignore the interference of additive noise during the optimization process, and are still insufficient in processing spatial details and inter-pixel information perception, resulting in poor restoration results. In this paper, we rethink the coarse-to-fine scheme and design a multi-scale structured atmospheric turbulence blind deblurring network (BATDNet), which consists of a shallow feature extraction module (SFE), an asymmetric feature fusion module (AFF), and a basic block called NFSR.SFE uses a simple structure to extract coarse-grained features at a shallow level to suppress noise and capture rich feature maps. The AFF module fuses multi-scale information of different granularities to effectively help the flow of feature information. NFSR, as a basic block, reduces the complexity of the model, enhances the importance of feature channels and space, and expands location-aware information. The backbone uses an encoder-decoder architecture to learn contextual information, which is then combined with a high-resolution branch that preserves spatial details to ensure high-quality image recovery. Our method demonstrates excellent performance on multiple simulated data scenarios and real atmospheric turbulence datasets.