Anomaly detection of small targets in PBF powder spreading process based on mask fusion

Enhancing the accuracy of anomaly detection in powder bed fusion, particularly for small-scale defects, remains an open challenge in additive manufacturing. In this work, we present a dataset of 7,116 layer-wise images collected from industrial PBF processes, with manual annotations covering five representative anomaly types. To address the difficulty of detecting such subtle defects, we design a detection framework that combines YOLOv8 with a background modeling module. The module incorporates hierarchical updates and a noise-aware regularization strategy, which together help to suppress random texture artifacts while preserving temporal priors. Furthermore, a feature fusion mechanism is introduced to enhance small-target localization by leveraging complementary. Finally, a dehaze inspired component is incorporated into the network, resulting in significant performance increases on the metrics tested. On the spread dataset, the proposed approach achieves an mAP50 of 0.88. This corresponds to improvements of up to 60.7% and 34.5% for small targets compared with standard baselines. Overall, the results highlight the importance of background-aware fusion strategies for reliable anomaly detection in PBF.