Synthesizing Data via Generative AI for Crack Detection in Metal Castings

Sammanfattning

Industrial foundries require reliable inspection systems for detecting small surface cracks. However, current supervised deep learning detection models require large annotated datasets, which are difficult to collect due to the rarity of defects and the time-intensive manual labour required. This thesis, conducted in collaboration with RISE and the SMYG (Smart image analysis for surface defects on cast components) project, investigates whether generative artificial intelligence can be used to synthesize surface-crack training data for a detection model. A small dataset of large, high-resolution images was collected at the Combi Wear Parts foundry and manually annotated. Since the cracks are difficult to observe with the naked eye, magnetic particle inspection was used to make them visible under UV illumination. The proposed pipeline uses a custom, mask-conditioned latent diffusion model, named DefectFill. This model generates realistic UV cracks on defect-free metal surfaces. The generated data were evaluated by training segmentation models and testing them on real cracks. An ablation study found that a hybrid dataset with 10% real data and 90% synthetic data achieved the highest performance, with mAP@50 = 0.363 and F1 = 0.53, compared with mAP@50 = 0.125 and F1 = 0.23 for the 100% real dataset. Tuning the hyperparameters using a genetic algorithm improved this to mAP@50 = 0.395 and F1 = 0.58. Furthermore, by testing the semantic reasoning of a large language model, a different set of hyperparameters achieved mAP@50 = 0.400 and F1 = 0.81, detecting 11 of 13 real cracks with only three false positives. Overall, the results demonstrate that synthetic data can improve detection rates in inspection systems when real data are limited.