基于改进YOLOv5模型的烟支外观缺陷检测  

Cigarette Appearance Defect Detection Based on Improved YOLOv5s Model

作　　者：钟林富 何逸波 舒梦 屠建飞[1] 王钢明[1] Zhong Linfu;He Yibo;Shu Meng

机构地区：[1]宁波大学先进储能技术与装备研究院,浙江宁波315211 [2]浙江中烟宁波卷烟厂,浙江宁波315211

出　　处：《机械制造》2025年第3期70-74,共5页Machinery

基　　金：国家自然基金资助项目(编号:22078164);宁波市重点研发项目(编号:2023Z153,2023Z061)。

摘　　要：烟支生产中常出现污渍、划痕、褶皱等外观缺陷,严重影响产品质量。针对烟支外观缺陷检测需求,提出一种改进YOLOv5s模型,可以提高烟支外观缺陷检测的准确性和效率。在主干网络的第一个卷积层中引入注意力机制,提高精准度。通过结合深度可分离卷积和通道注意力机制,对空间金字塔池化结构进行改造,提高鲁棒性。引入GhostConv和C3Chost,减少网络参数和计算量。试验结果表明,与标准YOLOv5s模型相比,改进YOLOv5s模型召回率提高3.7个百分点,重叠度阅值取0.5时平均精度均值提高3.2个百分点,浮点运算数减少31%,实现了检测性能的有效提高。Appearance defects such as stain,scratch,and wrinkle often occur in cigarette production,seriously affecting product quality.An improved YOLOv5s model was proposed to meet the demand for cigarette appearance defect detection,which can improve the accuracy and efficiency of cigarette appearance defect detection.The attention mechanism is introduced in the first convolutional layer of the Backbone to improve accuracy.By combining depthwise separable convolution and channel attention mechanism,the spatial pyramid pooling structure is modified to improve robustness.The GhostConv and C3Ghost were introduced to reduce network parameters and computational complexity.The experimental result shows that compared with the standard YOLOv5s model,the improved YOLOv5s model has a 3.7 percentage point increase in recall rate,a 3.2 percentage point increase in mean average precision when the overlap threshold is set to 0.5,and a 31%reduction in floating-point operations,and achieves effective improvement in detection performance.

关 键 词：卷烟 外观 缺陷 检测 模型 

分 类 号：TS47[农业科学—烟草工业]