HGNetV2-YOLO: An Efficient and Lightweight Framework for Mechanical Component Detection in Industrial Scenes
TL;DR
A lightweight YOLO-style detector that integrates a PP-HGNetV2 tiny backbone, an enhanced normalization-based attention module (ImNAM), and an improved complete intersection-over-union loss (ImCIoU) shows strong potential for intelligent industrial inspection on resource-constrained platforms, subject to further hardware-level deployment verification.
Abstract
Mechanical component detection in industrial scenes is challenged by cluttered backgrounds, large-scale variation, specular reflection, high inter-class similarity, and class imbalance. To address the above problems, this paper proposes a lightweight YOLO-style detector that integrates a PP-HGNetV2 tiny backbone, an enhanced normalization-based attention module (ImNAM), and an improved complete intersection-over-union loss (ImCIoU). The HGNetV2 backbone enhances hierarchical multi-scale feature extraction and keeps the deployable computational complexity low. ImNAM has been modified to enhance discriminative representation by introducing dual-statistics channel weighting, orthogonal edge-aware spatial modeling and bipolar adaptive residual gating. ImCIoU enhances the accuracy of localization by combining quality-aware box scaling, scale-sensitive modulation and dynamic IoU-guided weighting. A class-balancing augmentation pipeline was applied to the four-category industrial dataset of Bearing, Bolt, Gear and Nut. All experimental results are reported as the mean ± standard deviation of five independent two-tailed training runs with different random seeds, and statistical significance is verified by paired t-tests (p < 0.05) with Bonferroni correction for multiple comparisons. Experimental results show that the proposed method achieves 90.82 ± 0.35% mean average precision (mAP@0.5), 91.95 ± 0.42% precision, and 82.98 ± 0.51% recall, outperforming nine mainstream lightweight detectors, including the latest YOLOv12n (2025) and RT-DETR-tiny. Extended evaluation on mAP@0.5:0.95, per-class AP and F1 score further confirms the advantages in localization accuracy and classification performance. Ablation studies confirm that the HGNetV2 family backbone provides the largest performance gain, while the improved attention mechanism and regression loss further enhance localization accuracy and robustness. With only 4.44 M parameters and 9.96GFLOPs, the proposed detector has achieved a good accuracy–efficiency trade-off and shows strong potential for intelligent industrial inspection on resource-constrained platforms, subject to further hardware-level deployment verification.