基于深度特征融合的小尺寸杆端关节轴承检测方法  

Detection Method of Small-Size Rod-End Joint Bearing Based on Deep Feature Fusion

作　　者：叶瑞峰 彭晋民 宋衍聪 赵文婷 YE Rui-feng;PENG Jin-min;SONG Yan-cong;ZHAO Wen-ting(Fujian Provincial Key Laboratory of Intelligent Machining Technology and Equipment(Fujian University of Technology),Fuzhou 350118;Fujian University of Science and Technology,Mechanical and School of automotive Engineering,Fuzhou 350118;Xili Subdistrict Office,Nanshan District,Shenzhen 518055)

机构地区：[1]福建省智能加工技术及装备重点实验室(福建理工大学),福建福州350118 [2]福建理工大学机械与汽车工程学院,福建福州350118 [3]深圳市南山区西丽街道办事处,广东深圳518055

出　　处：《制造业自动化》2025年第1期45-52,共8页Manufacturing Automation

基　　金：福建省高校产学合作项目(2020H6028);2022年福建省省级科技创新重点项目(2022G02007);福建省高等学校科技创新团队(闽教科[2020]12号)。

摘　　要：运用机器视觉检测密集、小尺寸杆端关节轴承存在特征信息少以及滑动球面的高可变性,导致识别不精确进而影响生产效率的问题;由此,提出一种深度学习目标检测算法模型,首先为了使网络携带更多语义信息,引入Space-to-depth Convolution(SPD-Conv)无步长卷积模块改进Backbone网络;提出Multilevel Feature Fused SPD(MFSPD)深度特征融合模块重新设计Neck网络,提升对小尺寸目标的特征信息提取能力与检测精度;在Head网络增加一个P4小尺寸检测头,使用加权K-means算法在数据集上获取先验框,增大先验框与特征图层匹配度,加快模型收敛速度;接着引入置信度传播聚类分析算法Confidence Propagation Cluster(CP-Cluster)做后处理,优化预测框置信度与检测速度;然后最后在自制数据集、TLESS数据集以及COCO数据集上评估算法性能,该目标检测算法在自制数据集与T-LESS数据集上mAP@.5达到96.8%与93.6%,在COCO数据集上mAP为55.7%,实验结果表明算法检测精度与特征信息提取能力效果显著。Detection of small-sized rod-end joint bearings in dense arrangements of small-sized rod-end joint bearings via machine vision faces challenges due to the lack of feature information and high variability of the spherical surfaces,leading to inaccurate identification which affects production efficiency.To address this,a deep learning based object detection algorithm is proposed.First,to enrich the semantic information in the network,a Space-to-depth Convolution(SPD-Conv)module without striding is introduced to improve to the Backbone network.A Multi-level Feature Fused SPD(MFSPD)module is then proposed to redesign the Neck network for enhanced feature extraction and detection accuracy on small objects.In the Head network,a P4 small objects detection branch with prior boxes generated by a weighted k-means algorithm on the dataset is added to increase feature-prior box matching and accelerate convergence.Next,a Confidence Propagation Cluster(CP-Cluster)post-processing method is applied to optimize predicted box confidence and detection speed.Finally,the algorithm performance on custom,T-LESS and COCO datasets is evaluated.The proposed detector achieves 96.8%and 93.6%mAP@.5 on custom and T-LESS datasets respectively, and 55.7% mAP on COCO, demonstrating remarkable improve‐ments in detection accuracy and feature extraction capability.

关 键 词：小尺寸 杆端关节轴承 MFSPD CP-Cluster 加权K-means 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]