基于BN-YOLOv5的轻量级齿轮表面缺陷检测方法  

Research on the Lightweight Gear Surface Defect Detection Algorithm Based on BN-YOLOv5

作　　者：赵小惠[1] 张智杰 胡胜 郇凯旋 刘磊 蒲军平 Zhao Xiaohui;Zhang Zhijie;Hu Sheng;Huan Kaixuan;Liu Lei;Pu Junping(School of Mechanical and Electrical Engineering,Xi'an Polytechnic University,Xi'an 710048,China)

机构地区：[1]西安工程大学机电工程学院,陕西西安710048

出　　处：《机械传动》2024年第5期145-151,共7页Journal of Mechanical Transmission

基　　金：国家自然科学基金项目(72001166);陕西省科技计划项目(2022JQ-721);陕西省社会科学界联合会项目(20ZD195-59)。

摘　　要：齿轮表面的缺陷检测是齿轮生产制造过程中相当重要的工序。为了提高齿轮表面缺陷检测的精度,提出了一种基于改进YOLOv5的算法检测模型BN-YOLOv5。首先,将加权双向特征金字塔网络结构嵌入到颈部网络结构中,强化了网络对不同特征的提取能力;其次,引入轻量级的基于标准化的注意力模块(Normalization-based Attention Module,NAM),将其与加权双向特征金字塔网络结构相结合,以更加有效地融合高层与低层的特征信息;最后,采用深度可分离卷积模块替换网络结构中所有的卷积层,使网络模型更加轻量化。实验结果显示,改进后的算法模型平均精度均值可达到98.5%,检测速度达到66 FPS/s,模型大小为9.69 MB,有效降低了模型的占用内存,可满足在小型移动设备上实时检测齿轮表面缺陷的任务要求。A pretty crucial step in the manufacturing of gears is the defect detection on gear surfaces.An algorithmic detection model called BN-YOLOv5 which is based on an improved YOLOv5 is proposed in order to increase the accuracy of gear surface defect detection.Firstly,the technique strengthens the network's capacity to extract various features by embedding the weighted bidirectional feature pyramid network structure into the neck network structure.Secondly,a compact focus mechanism module,normalization-based attention module(NAM)is presented to its weighted bidirectional feature pyramid network structure which can more rapidly and efficiently fuse the feature information of higher and lower layers.Finally,the depth separable convolution mod-ule is used to replace every convolutional layer in the network structure,thereby lightening the network model.The experimental findings demonstrate that the enhanced algorithm model can achieve an average accuracy of 98.5%,a detection speed of 66 frames per second,and a modelling size of 9.69 MB,which effectively reduces the memory footprint of the model,and enables the task of real-time inspection of gear surface defects on small mobile devices.

关 键 词：齿轮表面 缺陷检测 YOLOv5 轻量级 NAM注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]