DDEG: Addressing data scarcity in additive manufacturing defect detection through joint generation of CT images and defect annotations

Deep learning offers significant advantages in computer vision but relies heavily on large annotated datasets. The application of deep learning in industrial computed tomography (CT) inspection of additive manufacturing (AM) components is severely constrained by two critical challenges: the scarcity of representative defect samples and the high cost of manual annotation processes. This paper introduces Defect Distribution and Edge-guided Generation (DDEG), a novel two-stage framework that simultaneously generates high-quality CT images and corresponding defect annotations using limited real samples. The first stage employs a Defect Distribution Generator (DDG), which uses a transfer learning-based latent diffusion model to learn defect morphology and spatial distribution patterns from real samples, generating defect annotations and edge images that match real statistical patterns. The second stage implements an Edge-guided Image Generator (EIG) with Low-rank Control Adapters (LoCA) that optimizes the condition encoding mechanism, enabling high-fidelity image synthesis through parameter-efficient feature alignment. Experiments show that DDEG-generated images achieve a 46.2% reduction in FID and a 4.3% improvement in NIMA compared to the best alternative method, a 25% improvement in diversity, and a 15.6% enhancement in annotation consistency. Defect detection models trained with our generated data show improvements of 3.5 percentage points in precision, 22.2 percentage points in recall, and 12.1 percentage points in mean Average Precision. This method provides an economical and efficient data augmentation solution for industrial inspection scenarios with limited samples.