Image Restoration Methods for Imaging Through Atmospheric Turbulence

Zhiyuan Mao, Purdue University


The performance of long-range imaging systems often suffers due to the presence of atmospheric turbulence. One way to alleviate the degradation caused by atmospheric turbulence is to apply post-processing mitigation algorithms, where a high-quality frame is reconstructed from a single degraded image or a sequence of degraded frames. The image processing algorithms for atmospheric turbulence mitigation have been studied for decades, yet some critical problems remain open. This dissertation addresses the problem of image reconstruction through atmospheric turbulence from three unique perspectives: 1. Reconstruction with the presence of moving objects. By re-designing the reference frame extraction step, our method can keep the moving object in the reference frame, and thus the conventional pipeline can be extended to dynamic scenes. We also introduce a robust way to perform lucky region fusion. In the end, we propose a physics-constrained prior model of the point spread function for blind deconvolution. Our reconstruction method achieves state-of-the-art performance among classic optimization-based methods in both static and dynamic scene cases. 2. A fast turbulence simulation scheme for generating large-scale datasets. We propose the Phase-to-Space transform, a lightweight neural network module that transforms the Zernike Polynomial weights to learned dictionary weights to synthesize the point spread functions efficiently. Our approach is 1000 times faster than the classic split-step simulation methods on GPU while keeping the essential turbulence statistics. 3. We propose a physics-inspired transformer model for single-frame image reconstruction through atmospheric turbulence. The proposed network utilizes the power of transformer blocks to jointly extract a dynamical turbulence distortion map and restore a turbulence-free image. We also present a real-world dataset for better evaluation of turbulence mitigation algorithms




Chan, Purdue University.

Subject Area

Optics|Electrical engineering|Remote sensing|Atmospheric sciences

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server