Deep learning-driven x-ray digital tomosynthesis (DT) imaging for aerospace composite inspection

Abstract

The structural integrity of aerospace-grade Glass Fiber Reinforced Polymer (GFRP) composites is critical, yet conventional non-destructive testing (NDT) methods often struggle to detect subsurface defects reliably due to poor signal-to-noise ratios, low contrast, and complex internal structures. To address these limitations, this study proposes a novel AI-driven framework that integrates low-power X-ray Digital Tomosynthesis (DT) imaging with state-of-the-art deep learning models for defect segmentation in composite materials. Specifically, two state-of-the-art instance segmentation models, YOLOv8 (You Only Look Once, version 8) and Detectron2, are employed to automatically segment flaws in the DT images of the composite specimens. A dedicated dataset of low-power X-ray DT scans of GFRP composite specimens with annotated defects was curated for training and evaluation. The segmentation performance of each model was quantitatively evaluated using metrics such as the Dice similarity coefficient and Intersection-over-Union (IoU), along with inference time measurements. Experimental results demonstrate that YOLOv8 processes images significantly faster (~6.9 ms per image) than Detectron2 (~10.3 ms), enabling near real-time analysis. Conversely, Detectron2 achieves a higher segmentation accuracy (Dice ~86% versus ~74% for YOLOv8), underscoring the trade-off between computational efficiency and segmentation precision. These findings validate the potential of combining low-power DT imaging with deep learning for high-fidelity defect identification, substantially improving the prospects for near real-time composite inspection. Future work will focus on further model optimization (e.g., via quantization and pruning) and the integration of this framework with autonomous robotic inspection systems, thereby extending the capabilities of AI-driven NDT in aerospace applications.