Airport-FOD3S: A Three-Stage Detection-Driven Framework for Realistic Foreign Object Debris Synthesis

Area of Science:

• Computer Vision
• Artificial Intelligence
• Aerospace Engineering

Background:

• Traditional Foreign Object Debris (FOD) detection methods struggle with large datasets and low accuracy.
• Effective FOD detection is critical for preventing accidents in aviation and other industries.

Purpose of the Study:

• To develop advanced data augmentation techniques for FOD detection.
• To improve the accuracy and robustness of FOD detection algorithms.

Main Methods:

• Utilized generative adversarial networks (GANs) and diffusion models for image data augmentation.
• Proposed a three-stage image blending method incorporating size transformation, seamless processing, and style transfer.
• Evaluated image quality using metrics like SSIM, PSNR, and Depthanything.
• Tested object detection models (Faster R-CNN, YOLOv8, YOLOv11) with a similarity distance strategy (SimD).

Main Results:

• Generated realistic FOD images under diverse environmental conditions.
• The proposed three-stage blending method achieved superior image quality (SSIM=0.99, PSNR=45 dB).
• YOLOv11 with SimD, trained on augmented data, reached a mean Average Precision (mAP) of 86.95%.

Conclusions:

• Data augmentation and the SimD strategy significantly enhance FOD detection accuracy.
• The developed methods offer a promising solution for overcoming limitations in traditional FOD detection.