基于YOLOv5⁃DCN的织物疵点检测  

Fabric defect detection based on YOLOv5⁃DCN

作　　者：胡越杰 蒋高明[1,2] HU Yuejie;JIANG Gaoming(Jiangnan University,Wuxi,214122,China;Engineering Research Center for Knitting Technology,Ministry of Education,Wuxi,214122,China)

机构地区：[1]江南大学,江苏无锡214122 [2]针织技术教育部工程研究中心,江苏无锡214122

出　　处：《棉纺织技术》2023年第3期8-14,共7页Cotton Textile Technology

基　　金：国家自然科学基金项目(61772238);泰山产业领军人才项目(tscy20180224)。

摘　　要：探讨基于改进的YOLOv5织物疵点检测算法。为了增强网络模型的特征提取能力,在主干特征提取网络的卷积层中引入可变形卷积,并设计ResDCN模块用于取代原网络中的残差单元模块。针对织物疵点边缘轮廓不规则、目标区域占比小等特征,将原边框回归损失函数改进为Focal EIoU损失函数。该模型采用DyReLU激活函数,将动态卷积核与动态激活函数相结合,显著提高了织物疵点检测的准确性。试验结果表明:与YOLOv5s模型相比,YOLOv5⁃DCN模型的精准率、召回率和mAP@0.5值分别提升了4.99个百分点、2.26个百分点和2.49个百分点。认为:基于YOLOv5⁃DCN的织物疵点检测算法可为复杂环境下织物疵点的高效识别提供可靠的技术支持。Fabric defect detection algorithm based on improved YOLOv5 was discussed.In order to enhance the feature extraction capability of the network model,deformable convolution was introduced into the convolutional layer of the backbone feature extraction network,the ResDCN module was designed to replace the residual unit module in the original network.Aimed at the characteristics of irregular edge contour of fabric defects and small proportion of target area,the original bounding box regression loss function was improved to Focal EIoU loss function.DyReLU activation function was adopted in this model.Dynamic convolution kernel and dynamic activation function were combined.The accuracy of fabric defect detection was significantly improved.The experimental results showed that the precision,recall and mAP@0.5 of the YOLOv5⁃DCN model were improved by 4.99 percentage points,2.26 percentage points and 2.49 percentage points respectively compared with the YOLOv5s model.It is considered that the fabric defect detection method based on YOLOv5⁃DCN can provide reliable technical support for efficient identification of fabric defects in complex environments.

关 键 词：疵点检测 YOLOv5模型 卷积神经网络 可变形卷积 DyReLU激活函数 

分 类 号：TS941.26[轻工技术与工程—服装设计与工程]