Active learning for industrial defect detection: a study on hybrid sampling strategies

In modern industrial environments, ensuring the quality of manufactured components is critical, particularly when dealing with reflective surfaces that hinder conventional inspection techniques. Although deep learning-based methods offer robust solutions for visual defect detection, their performance often hinges on the availability of substantial annotated datasets. In industrial scenarios, labeling such datasets is costly and time-consuming. This study investigates applying sample selection techniques to reduce annotation efforts for porosity detection on machined aluminium parts. Several selection strategies were evaluated using a real-world dataset composed of high-resolution images, including uncertainty, diversity, random-based criteria, and hybrid combinations. The best-performing strategy, which combined entropy-based uncertainty, spatial diversity, and random-based, achieved an F1-score of 86.70% and a recall of 82.99% after ten iterations using only 2,400 annotated images, corresponding to 66.67% of the active learning pool. Although the fully supervised model achieved an F1-score of 88.84% and a recall of 86.30%, the proposed approach proved a competitive alternative. These results demonstrate that selective data annotation can significantly reduce labeling effort while maintaining reliable performance in defect detection, even under the challenging conditions posed by reflective industrial parts.