Automated defect recognition (ADR) using non-destructive testing (NDT) is increasingly important for quality assurance in additively manufactured components. However, deep-learning-based ADR remains limited by scarce defect samples, severe class imbalance, and costly pixel-level annotations. This paper presents a unified physics-inspired diffusion framework for synthetic defect generation across digital radiography, X-ray computed tomography, and active thermography. A defect-first strategy samples parametric defect geometries and embeds them into modality-specific forward models based on Beer–Lambert attenuation and heat diffusion. A conditional denoising diffusion probabilistic model then refines the physics-rendered samples to reproduce realistic sensor appearance while preserving label fidelity through projection-consistency and thermal partial differential equation residual losses. For segmentation, pseudo-masks are generated automatically using adaptive thresholding and Canny edge detection, eliminating manual annotation. A U-Net encoder–decoder is trained using combined binary cross-entropy and Dice losses and evaluated under low-data, noise-stress, and out-of-distribution defect-shift conditions. The reproducible benchmark includes 10,000 radiography, 5,000 CT, and 5,000 thermography samples at 256 × 256 resolution, with external validation on GDXray weld and Bridge Cracks datasets. Physics-constrained diffusion augmentation improves segmentation Dice by 4.2–11.7 percentage points and classification F1 by 3.8–9.3 percentage points over real-only training.
Baskar et al. (Wed,) studied this question.