基于YOLOv8n改进的织物疵点检测算法  

Improved fabric defect detection algorithm based on YOLOv8n

作　　者：刘伟宏 李敏[1,2,3] 朱萍 崔树芹[1] 颜小运[1] LIU Weihong;LI Min;ZHU Ping;CUI Shuqin;YAN Xiaoyun(Wuhan Textile University,Wuhan,430200,China;Engineering Research Center of Hubei Province for Clothing Information,Wuhan,430200,China;Hubei Engineering Research Center of Intelligent Textile and Fashion,Wuhan,430200,China)

机构地区：[1]武汉纺织大学,湖北武汉430200 [2]湖北省服装信息化工程技术研究中心,湖北武汉430200 [3]纺织服装智能化湖北省工程研究中心,湖北武汉430200

出　　处：《棉纺织技术》2024年第10期19-25,共7页Cotton Textile Technology

基　　金：湖北省教育厅科学研究计划重点项目(D20211701);湖北省自然科学基金(2022CFC074)。

摘　　要：为了解决织物疵点检测中小目标疵点难以检测的问题,提出了一种基于改进YOLOv8n算法的织物疵点检测系统。首先,在特征融合部分,采用了兼顾速度和精度的GSConv替代原有的卷积核,并引入了Slim⁃neck特征融合模块,使每个特征层能够同时考虑深层特征的语义信息和浅层特征的细节信息,提高了对小目标的特征响应,同时简化了模型并降低了计算复杂度。其次,设计了用于检测小疵点目标的检测层P2,增强了模型对小疵点目标的检测能力,使其更适用于织物疵点检测任务。最后,采用指数滑动样本加权函数(EMA⁃SlideLoss)替代了交叉熵损失函数,以增强模型的类别分类能力,提高训练的稳定性。试验结果表明:在检测20类疵点时,相较于YOLOv8n模型,该研究方法在mAP@0.5方面提高了0.142,同时实现了47.4帧/s的检测速度。改进的YOLOv8n模型对网络的性能提升是有效的。To address the difficulty of detecting middle and small-sized defect targets in fabric inspection,an improved fabric defect detection system based on YOLOv8n algorithm was proposed.Firstly,in the feature fusion part,a speed and accuracy balanced GSConv was used to replace the original convolution kernel,and a Slim-neck feature fusion module was introduced,which allowed each feature layer to consider both the semantic information of deep features and the detail information of shallow features.The feature response to small targets was enhanced.Meanwhile,the model was simplified and the computational complexity was reduced.Secondly,a detection layer P2 was designed for detecting small defect targets,enhancing the model’s ability to detect small defect targets and making it more suitable for fabric defect detection tasks.Finally,an Exponential Moving Average Sample Weighting Function(EMA-SlideLoss)was used to replace the cross-entropy loss function to enhance the model’s category classification ability and improve training stability.Experimental results showed when detecting 20 types of defects,compared with YOLOv8n,mAP@0.5 of this method was improved by 0.142 and achieved a detection speed of 47.4frame/s.The improved YOLOv8n model was effective to improve the performance of the network.

关 键 词：织物疵点 YOLOv8n算法 Slim⁃neck EMA⁃Slideloss GSConv 

分 类 号：TS101.8[轻工技术与工程—纺织工程]