DPAC: a dual-path domain adaptive fasteners detection method based on adaptive cutmix strategy

Object detection technology can assist in real-time monitoring of wrong and missing assemblies in the assembly process. However, the current diverse production assembly requires massive labeled data for supervised object detection, which is time-consuming and labor-intensive. To this end, this study focuses on common fasteners in workshops, such as bolts and nuts, and proposes a dual-path domain adaptive detection method based on an adaptive cutmix strategy (DPAC). The method achieves object detection using labeled source data and unlabeled target data as inputs. To achieve higher accuracy in fastener assembly detection, the method uses Mamba + YOLOv5 as the backbone and integrates a data augmentation scheme that combines multiple image transformations. Then, an adaptive cutmix strategy is developed to ensure the integrity of the mixed image labels and address the label occlusion issue in the mixing process. Following this, a dual-path strategy for source and target data is designed, which contributes to a stronger generalization ability of the training model and a higher accuracy of fastener object detection. Finally, a fastener detection dataset is constructed, and on this dataset, the effectiveness of the proposed strategy and methods is validated. Experimental results show that the mean average precision of the proposed DPAC framework reaches 80.4 % for the detection of workshop fasteners in assembly monitoring and 50.7 % on general datasets (Sim10k-Cityscapes). These results validate the applicability of our method to fastener assembly monitoring in workshops.