基于改进YOLOv5s的皮革面料疵点检测  

Leather fabrics defect detection based on an improved YOLOv5s

作　　者：汪佳超 陈敏之 WANG Jiachao;CHEN Minzhi(School of Fashion Design and Engineering Department,Zhejiang Sci-Tech University,Hangzhou 310018,China;School of International Education,Zhejiang Sci-Tech University,Hangzhou 310018,China)

机构地区：[1]浙江理工大学服装学院,浙江杭州310018 [2]浙江理工大学国际教育学院,浙江杭州310018

出　　处：《染整技术》2024年第12期35-41,共7页Textile Dyeing and Finishing Journal

摘　　要：基于机器学习的疵点检测方法已经在纺织面料领域得到广泛应用,但仍存在检测精度不高、检测速度不快的问题,无法达到工业实时检测的目的。针对这些问题提出了一种基于改进YOLOv5s的皮革面料疵点检测算法,命名为G3-YOLOv5。在Backbone中引入了Ghostconv,同时还将GSConv轻量卷积结构整合到Neck中,以降低模型的参数量,提高模型的检测速度。最后加入全局注意力机制(Global Attention Mechanism,GAM),提高模型对不同尺寸疵点的感知能力,在保证检测速度的前提下,提高模型的检测精度。试验数据表明,在自建数据集上,G3-YOLOv5算法的mAP值比传统的YOLOv5s高出7.61%,达到了86.2%;在CPU上的检测速度(FPS)为2.68 Hz,比原始YOLOv5s快0.31 Hz,基本上满足了皮革面料疵点的检测需求。Defect detection methods based on machine learning have been widely used in the field of fabrics,but there are still the problem of low detection accuracy and fast detection speed,which can’t reach the purpose of industrial real-time detection.An improved YOLOv5s based defect detection algorithm for leather fabrics,named G3-YOLOv5,is proposed to address these issues.Ghostconv is introduced in Backbone,and a lightweight convolutional structure called GSConv is also integrated into Neck to reduce the number of parameters in the model and improve the detection efficiency.Finally,the GAM attention mechanism is added to improve the model’s ability to perceive defects of different sizes,which improves the model’s detection accuracy while ensuring the detection speed.According to experimental findings,the mAP value(86.2%)of the G3-YOLOv5 algorithm is 7.61%higher than that of YOLOv5s on the self-built dataset.The detection speed(FPS)on CUP is 2.68 Hz,which is 0.31 Hz faster than that of the original YOLOv5s,which basically meets the needs of detecting defects in leather fabric.

关 键 词：深度学习 皮革面料疵点 轻量化 注意力机制 

分 类 号：TS941.8[轻工技术与工程—服装设计与工程]