Removing Atmospheric Turbulence via Deep Adversarial Learning.

Area of Science:

• Computer Vision
• Deep Learning
• Image Restoration

Background:

• Atmospheric turbulence causes complex image distortions, including geometric distortion and blur.
• Single-stage deep networks struggle to mitigate the mixed distortions from atmospheric turbulence.
• Existing methods often require prior knowledge or multiple images, limiting their practical application.

Purpose of the Study:

• To develop an advanced deep learning model for robust image restoration under atmospheric turbulence.
• To address the limitations of single-stage networks in handling complex atmospheric distortions.
• To create a versatile restoration model applicable without prior turbulence condition knowledge or multi-image fusion.

Main Methods:

• Proposed a two-stage deep adversarial network (DT-GAN+ and DTD-GAN+) for image restoration.
• The first stage addresses geometric distortion, while the second stage minimizes image blur.
• Incorporated channel attention and a novel sub-pixel mechanism to enhance restoration at finer levels.
• Synthesized turbulent image datasets and curated a natural turbulent dataset from YouTube for training and validation.

Main Results:

• The proposed DT-GAN+ and DTD-GAN+ models significantly outperform state-of-the-art image-to-image translation and baseline restoration models.
• Restored images showed marked improvement in downstream tasks like classification, pose estimation, and depth estimation.
• The model demonstrated generalisability on natural turbulent image datasets.

Conclusions:

• A two-stage adversarial network with channel attention and sub-pixel mechanisms effectively restores images degraded by atmospheric turbulence.
• The proposed method offers a superior solution for atmospheric turbulence mitigation compared to existing approaches.
• The enhanced image quality translates to significant performance gains in various computer vision applications.