基于YOLOv5n的轻量级织物疵点检测算法  

Lightweight fabric defect detection algorithm based on YOLOv5n

作　　者：李洋 李敏[2,3] 黄政 董雄伟[1] 朱立成[1] LI Yang;LI Min;HUANG Zheng;DONG Xiongwei;ZHU Licheng(National Local Joint Engineering Laboratory for Advanced Textile Processing and Clean Production,Wuhan Textile University,Wuhan,Hubei 430200,China;School of Computer Science and Artificial Intelligence,Wuhan Textile University,Wuhan,Hubei 430200,China;Hubei Provincial Engineering Research Center for Intelligent Textile and Fashion,Wuhan Textile University,Wuhan,Hubei 430200,China;Best Textile Co.,Ltd.,Yichang,Hubei 443202,China)

机构地区：[1]武汉纺织大学先进纺纱织造及清洁生产国家地方联合工程实验室,湖北武汉430200 [2]武汉纺织大学计算机与人工智能学院,湖北武汉430200 [3]武汉纺织大学纺织服装智能化湖北省工程研究中心,湖北武汉430200 [4]毕生纺织有限公司,湖北宜昌443202

出　　处：《毛纺科技》2024年第5期87-97,共11页Wool Textile Journal

摘　　要：针对轻量级模型在检测织物疵点时精确率低的问题,在YOLOv5n的基础上提出一种上下文增强与混合感受野的织物疵点检测算法。首先,为主干网络设计了一种轻量扩张卷积空间金字塔模块,并将主干网络的下采样比增加至64,在增强上下文信息的同时提取更深层的语义信息,提高模型识别性能;其次,为颈部网络设计了一种混合感受野融合模块代替原C3模块并进行特征融合,提高极端长宽比目标的检测精度。实验表明:该算法在基于天池织物数据集上的IOU阈值为0.5时的平均精度均值mAP 50、精确率、召回率分别达到了93.1%、91.6%、89.1%,相较于原YOLOv5n算法分别提高了4.9%、7.3%、5.0%,且模型文件大小仅6.28 MB,更适用于织物疵点检测领域。To address the issue of low accuracy in detecting fabric defects using lightweight models,a fabric defect detection algorithm based on YOLOv5n was proposed,which combines context enhancement and mixed receptive fields.Firstly,a module of Spatial Pyramid for Lightweight Atrous was designed for the backbone network,and the down-sampling ratio of the backbone network was increased to 64,enhancing contextual information while extracting deeper semantic information to improve the recognition performance of the model;Secondly,a hybrid receptive field fusion module was designed for the neck network to replace the original C3 module for feature fusion,improving the detection accuracy of extreme aspect ratio targets.The experiment shows that the mAP 50,Precision,and Recall of this algorithm on the Tianchi fabric dataset have reached 93.1%,91.6%,and 89.1%,respectively.Compared with the original YOLOv5n,it has improved by 4.9%,7.3%,and 5.0%,and the model file size is only 6.28 MB,making it more suitable for the field of fabric defect detection.

关 键 词：疵点检测 深度学习 YOLOv5n 空间金字塔 感受野融合 

分 类 号：TS391.41[轻工技术与工程]