基于YOLOv5的瓶盖封装缺陷轻量化检测算法  

YOLOv5-Based Lightweight Algorithm for Detecting Bottle-Cap Packaging Defects

作　　者：赵磊[1,2,3] 矫立宽[1,2,3] 翟冉 李彬 许美叶[1,2,3] Zhao Lei;Jiao Likuan;Zhai Ran;Li Bin;Xu Meiye(Tianjin Key Laboratory for Advanced Mechatronic System Design and Intelligent Control,Tianjin University of Technology,Tianjin 300384,China;National Demonstration Center for Experimental Mechanical and Electrical Engineering Education(Tianjin University of Technology),Tianjin 300384,China;School of Mechanical Engineering,Tianjin University of Technology,Tianjin 300384,China)

机构地区：[1]天津理工大学天津市先进机电系统设计与智能控制重点实验室,天津300384 [2]机电工程国家级实验教学示范中心(天津理工大学),天津300384 [3]天津理工大学机械工程学院,天津300384

出　　处：《激光与光电子学进展》2023年第22期131-140,共10页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(51975412)。

摘　　要：为解决白酒瓶盖封装表面质量检测和算法参数庞大难部署的问题,对YOLOv5s进行改进并提出了更轻量化和高精度的SEGC-YOLO算法。首先,采用ShuffleNet V2替换原始骨干网络,有效简化参数,引入高效通道注意力机制增强骨干网络。再使用基于GhostNet改进的GhostConv和C3-Ghost模块增强颈部网络,减少颈部参数量。另外,使用CARAFE算子替代最近邻插值上采样算子,利用自适应内容感知的上采样预测核提升颈部网络的信息表达能力,进而提升检测精度。最后,训练应用Adam梯度优化器来提高检测精度。实验结果表明:所提SEGC-YOLO算法在不同交并比(IoU)阈值下的平均精度均值mAP@0.5为84.1%和mAP@0.5∶0.95为49.0%,分别优于原始YOLOv5s算法1.2个百分点和0.5个百分点,并且浮点运算数(FLOPs)比原始算法减少了69.94%、参数量减少了71.15%和模型文件大小减小了69.66%,更加精准和轻量化。所提SEGC-YOLO可以快速、精准地检测瓶盖表面缺陷,为相关领域快速缺陷检测和设备部署提供了数据和算法支持。With the aim to solve surface-quality detection of liquor bottle-cap packaging and the difficulty of deploying algorithms owing to large parameters,this study proposes a more lightweight and high-precision detection algorithm,named SEGC-YOLO,which is based on YOLOv5s.First,the ShuffleNet V2 is used to replace the original backbone network to effectively simplify the parameters,and the backbone network is enhanced using efficient channel attention mechanism.Next,the improved GhostConv and C3-Ghost modules,based on GhostNet,are used to improve the neck network and reduce the neck parameters.In addition,the CARAFE operator is introduced to replace the nearest neighbor interpolation upsampling operator.The upsampling prediction kernel with adaptive content awareness can improve the information-expression ability of the neck network and thereby the detection accuracy.The Adam gradient optimizer is used for training.Experimental results show that the proposed SEGC-YOLO algorithm achieves the mean accuracy precision mAP@0.5 of 84.1%and mAP@0.5∶0.95 of 49.0%at different intersection over union(IoU)thresholds,which are 1.2 and 0.5 percentage points higher than the original YOLOv5s algorithm,respectively.The overall floating-point operations(FLOPs),parameter volume,and model file size are also reduced by 69.94%,71.15%,and 69.66%,respectively,indicating higher accuracy and lighter weight compared with that of the original algorithm.Therefore,SEGCYOLO can quickly and accurately identify the surface defects of bottle caps,providing data and algorithm support for rapid detection and equipment deployment in related fields.

关 键 词：缺陷检测 轻量化算法 YOLOv5 ShuffleNet V2 GhostNet CARAFE算子 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]