Augmenting atmospheric turbulence effects on thermal-adapted deep object detection models

Area of Science:

• Computer Vision
• Atmospheric Optics
• Machine Learning

Background:

• Atmospheric turbulence distorts optical images, impacting object detection model performance.
• Variations in air's refractive index cause light scattering, leading to blurring and geometric distortions.

Purpose of the Study:

• To evaluate turbulence image augmentation for enhancing object detection models.
• To assess the robustness of deep learning models under atmospheric turbulence.

Main Methods:

• Employed three turbulence simulators (geometric, Zernike-based, P2S) to create turbulent datasets.
• Trained and tested RTMDet-x, DINO-4scale, and YOLOv8-x models with and without turbulence augmentation.
• Evaluated model performance on both turbulent and non-turbulent test sets.

Main Results:

• Turbulence-specific augmentations substantially improved object detection accuracy and robustness.
• Augmented models showed better performance on distorted images compared to non-augmented models.
• Performance gains were observed even on non-turbulent test data.

Conclusions:

• Turbulence image augmentation is an effective strategy for improving object detection in adverse atmospheric conditions.
• Deep learning models benefit from specialized augmentations to overcome turbulence-induced image degradation.
• The proposed augmentation techniques enhance model generalization capabilities.