改进YOLOv10n的紧固件检测算法

An Improved YOLOv10n Algorithm for Industrial Fastener Detection

摘要: 针对工业紧固件检测中存在的目标形态多样、堆叠遮挡及背景干扰等问题，提出一种改进的YOLOv10n-WGM检测算法。以M6螺母与M8螺丝为检测目标，基于Unity3D构建多场景仿真数据集，通过控制光照强度、方向、地面纹理与颜色等参数增强数据多样性，并模拟物理掉落过程生成堆叠目标图像，共采集600张图像按4:2划分训练集与验证集。在YOLOv10n基础上引入三方面改进：使用WIoU损失函数提升低质量样本处理能力，通过动态调节损失权重缓解锚框质量不平衡问题，优化梯度分配策略；将C2f模块替换为C2fCIB-LEGM结构，在提升局部特征提取能力的同时建模全局依赖；在检测头前融入CGAFusion模块，提升跨层级特征流动效率并保留空间细节特异性。实验结果表明，YOLOv10n-WGM相比原模型在mAP50、mAP50:95、精确度与召回率分别提升3.588%、1.551%、2.342%与4.295%，漏检率显著降低。该算法具有良好的检测精度与鲁棒性，适用于工业环境中的紧固件实时检测需求。

Abstract: To tackle challenges like varied shapes, occlusion, and background clutter in industrial fastener detection, we propose an enhanced YOLOv10n-WGM algorithm. Focusing on M6 nuts and M8 screws, a diverse synthetic dataset was generated using Unity3D by varying lighting, textures, and colors, and simulating physical stacking. A total of 600 images were collected and split into training and validation sets at a 4:2 ratio. The model incorporates three key improvements: 1) WIoU loss, which dynamically adjusts the loss weight to alleviate anchor box quality imbalance and optimize the gradient allocation strategy, for better handling of low-quality samples; 2) The C2f module is replaced by the C2fCIB-LEGM structure to enhance local feature extraction and model global dependencies; and 3) The CGAFusion module is integrated before the detection head to enhance cross-level feature flow efficiency and preserve spatial detail specificity. Experiments show significant gains: +3.588% mAP50, +1.551% mAP50:95, +2.342% precision, and +4.295% recall compared to the original model, with markedly lower missed detections. The method demonstrates high accuracy and robustness for real-time industrial fastener inspection.