SAM ‐ CompAug: Segment Anything‐Assisted Synthetic Data Augmentation for Industrial Part Detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Against the backdrop of the rapid advancement of intelligent manufacturing, part detection, as a core task within industrial vision systems, faces limitations due to the scarcity of high‐quality and fine‐grained datasets. To address this challenge, we propose a data augmentation method named SAM‐CompAug to enhance model robustness. The proposed methodology synergizes the Segment Anything Model (SAM) with bounding box prompts to generate high‐fidelity instance masks, while employing a minimum centre‐point distance matching strategy for precise label alignment. Furthermore, a Visibility‐aware Composite Placement Strategy is introduced to systematically generate synthetic samples with realistic occlusion patterns, overlapping configurations and scale variations in synthetic samples. Leveraging this augmentation method, we construct the Part detection Dataset (PartID), a synthetic dataset for industrial part detection. PartID encompasses 13 categories of industrial parts, comprising a total of 51,000 images and 903,164 annotated instances, thereby enriching the training data and enhancing network robustness for part detection in real‐world scenarios. Meanwhile, to mitigate the challenge of scale variance in part detection, a multi‐scale perception module, GLMF, is designed and integrated into the YOLOv8n and YOLOv12n frameworks, thereby enhancing the model's robustness and adaptability across varying part scales. Experimental results demonstrate that PartID delivers significant performance gains on the model and exhibits strong transferability in cross‐domain scenarios. Comparative experiments reveal that the strategy of pre‐training on synthetic data followed by fine‐tuning on real data achieves improvements of 1.7%, 0.3%, 0.4% and 0.9% in mAP50–95, mAP50, Precision and Recall, respectively, relative to the baseline algorithm. Our implementation is publicly available at: https://github.com/zhan‐wz/MAYOLO .