Composite X-ray defect detection via diffusion-based defect synthesis and physics-inspired fusion with YOLO-PCN

Automatic recognition of micro-defects in aerospace composite X-ray non-destructive testing (NDT) is difficult because defects are very small, weak in contrast, and often obscured by overlapping background structures. Limited annotations and class imbalance further reduce the reliability of conventional detectors. To address this problem, we propose a unified framework that improves training data, defect visibility, and small-defect localisation in one pipeline. A diffusion-based defect synthesis module expands scarce defect samples, a perception-aware pseudo-colour enhancement method improves weak-defect separability in high-bit-depth X-ray images, and a high-resolution detector improves the localisation of very small defects. Experiments on a real industrial composite X-ray dataset achieve a Precision of 0.954, a Recall of 0.939, and an mAP 50 of 0.963 with real-time inference, outperforming representative detectors under a unified benchmark. These results show that the proposed method provides an effective and deployment-oriented solution for micro-defect inspection in aerospace composites.