基于YOLOv5⁃Eff网络的织物疵点检测算法  

Fabric defect detection algorithm based on YOLOv5-Eff network

作　　者：石玉文 林富生[1,2,3] 宋志峰 余联庆 SHI Yuwen;LIN Fusheng;SONG Zhifeng;YU Lianqing(Wuhan Textile University,Wuhan,430200,China;Three Dimensional Textile Hubei Engineering Research Center of Hubei Province,Wuhan,430200,China;Hubei Key Laboratory of Digital Textile Equipment,Wuhan,430200,China)

机构地区：[1]武汉纺织大学,湖北武汉430200 [2]三维纺织湖北省工程研究中心,湖北武汉430200 [3]湖北省数字化纺织装备重点实验室,湖北武汉430200

出　　处：《棉纺织技术》2023年第12期20-25,共6页Cotton Textile Technology

摘　　要：针对目标检测精度低、小目标易漏检等问题,提出一种基于改进YOLOv5模型的检测算法。选取改进EfficientNet B1网络作为主干特征提取网络;引入ACmix注意力模块提高网络对小尺度目标的敏感度,降低噪声所带来的影响,解决小缺陷特征图在卷积操作中的失真情况;将SiLU与Swish激活函数结合,根据目标的数量和密度来动态调整阈值,提高算法灵活性。研究结果表明:相比于原始YOLOv5模型,改进后的YOLOv5算法的精确率、召回率和平均精度均值分别提升了4.33个百分点、2.11个百分点和4.32个百分点。该算法能准确识别织物疵点的整体特征,对于复杂场景下的疵点以及小目标疵点检测任务更为适用。Aiming at the problems of low target detection precision,miss detectioning easily for small targets and so on,a detection algorithm based on improved YOLOv5 model was proposed.The improved EfficientNet-B1 network was selected as the backbone feature extraction network.The attention module of ACMIX was introduced to improve the sensitivity of the network to small-scale targets,reduce the impact of noise and solve the distortion condition of small defect characteristic pattern during convolution operation.SiLU and Swish activation function were combined.Dynamic threshold value was adjusted and flexibility of the algorithm was improved according to the target quantity and density.The results showed that the precision,recall and mAP value of improved YOLOv5 algorithm were improved by 4.33 percentage points,2.11 percentage points and 4.32 percentage points respectively compared with the original YOLOv5 model.The algorithm could accurately identify the overall characteristics of fabric defects,and was more suitable for the detection of defects in complex scenes and small targets.

关 键 词：YOLOv5 EfficientNet 注意力模块 Swish动态激活函数 织物疵点 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]