Multi-Source Annotation Uncertainty Fusion Method for Aluminum Ingot Surface Defect Detection

Surface slag inclusions on aluminum ingots in high-temperature casting lines often appear in dense clusters. Under poor imaging conditions and with ambiguous defect boundaries, annotators frequently disagree on the location, number, and scale of defects, which limits the performance of conventional vision-based detection methods that rely on deterministic labels. To address this issue, this study proposes a Multi-source Annotation Uncertainty Fusion (MAUF) method. First, a spatial clustering and splitting strategy is introduced to decouple controversial labels from multiple sources into independent regions. Then, a controversy degree metric is defined to convert hard labels into soft supervision signals. Finally, a weighted loss function is designed to dynamically balance the model’s attention between high-controversy and low-controversy regions. In addition, to reduce the evaluation bias caused by traditional Intersection over Union (IoU)-based matching in this scenario, an area-based evaluation metric is developed. Experiments on a real aluminum ingot production dataset show that MAUF achieves an F1low of 48% at a confidence threshold of 0.3, improving by 9.84–26.32% over comparative methods in low-controversy regions while maintaining a moderate coverage rate of about 20% in high-controversy regions. The method also shows strong robustness across different confidence thresholds, with F1low remaining around 48%, whereas competing methods fluctuate more noticeably. Overall, MAUF provides a robust solution for defect detection in noisy industrial environments and offers a useful reference for handling multi-source annotation uncertainty.