Classification guided thick fog removal network for drone imaging: ClassifyCycle.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Drone imagery often suffers from non-uniform fog due to inhomogeneous distribution.
• Varying fog thickness in drone images presents significant challenges for clear visual data acquisition.

Purpose of the Study:

• To propose a novel classification-guided thick fog removal network for drone imaging.
• To enhance the reliability and reduce distortion in defogged drone images.

Main Methods:

• Introduced a classification module (ICLFn) to improve the reliability of the learning network.
• Incorporated a style migration module (ISMn) to minimize image distortions like hue artifacts and texture issues.
• Developed the ClassifyCycle network, which operates without requiring paired foggy and fog-free datasets.

Main Results:

• The ClassifyCycle network effectively removes thick fog from drone images.
• The method avoids common defogging issues such as overexposure, distortion, color deviation, and residual fog.
• Experimental results demonstrate superior performance compared to state-of-the-art algorithms on both synthetic and real-world drone images.

Conclusions:

• ClassifyCycle offers a robust solution for thick fog removal in drone imagery.
• The network's ability to function without paired data makes it a practical tool for real-world applications.
• The proposed approach significantly advances the quality of drone-based visual data in adverse weather conditions.