改进YOLOv5的布匹缺陷检测方法  

Method of fabric defect detection based on improved YOLOv5

作　　者：张凯旋 杜景林[2] ZHANG Kaixuan;DU Jinglin(College of Computer Science,Nanjing University of Information Science and Technology,Nanjing 210044,China;School of Artificial Intelligence,Nanjing University of Information Science and Technology,Nanjing 210044,China;NUIST-TianChang Research Institute,Tianchang 239300,China)

机构地区：[1]南京信息工程大学计算机学院,江苏南京210044 [2]南京信息工程大学人工智能学院,江苏南京210044 [3]南京信息工程大学天长研究生院,安徽天长239300

出　　处：《现代电子技术》2024年第20期109-117,共9页Modern Electronics Technique

基　　金：国家自然科学基金资助项目(41575155)。

摘　　要：现阶段布匹缺陷种类繁杂,且包含大量人眼难以辨别的小目标缺陷和长宽比极端不平衡缺陷,使得在复杂背景下的布匹缺陷检测成为一项艰巨任务。为此,提出一种改进YOLOv5的布匹缺陷检测方法。首先,在YOLOv5的C3模块中增加注意力机制NAM,设计为C3NAM模块,其可以抑制特征值中不显著的权重,在保持性能的同时进行高效计算;其次,采用一个新的CNN模块SPD-Conv,以解决大部分的布匹缺陷检测在分辨率较低或者瑕疵较小时性能迅速下降的问题;最后,在检测端引入新的损失函数Alpha-IoU,促进真实框和预测框的拟合,并提升对缺陷预测的准确性。实验结果表明:改进的YOLOv5网络模型较原YOLOv5网络模型mAP@0.5值提高了5.4%,mAP@0.5:0.95值提高了2.2%,且检测效果优于原网络模型和其他主流目标检测模型。There are various types of fabric defects,including a large number of small target defects that are difficult for the human eye to distinguish and extremely imbalanced aspect ratio defects,making fabric defect detection in complex backgrounds a daunting task.Therefore,a method of fabric defect detection based on improved YOLOv5 method is proposed.The C3NAM(normalization-aware mechanism)module is designed by adding attention mechanism NAM to the C3 module of YOLOv5,which can suppress the insignificant weights in the characteristic value and perform efficient computation while maintaining performance.A new CNN(convolutional neural networks)module SPD-Conv(space-to-depth-convolution),is adopted to solve the problem of rapid performance degradation in most fabric defect detection when the resolution is low or the defects are small.A new loss function Alpha-IoU is introduced in the detection side to facilitate the fitting of the real box and the prediction box,and improve the accuracy of the defect prediction.The experimental results show that,in comparison with the original YOLOv5 network model,the improved YOLOv5 network model has an increase of 5.4%mAP@0.5 and 2.2%mAP@0.5:0.95,and the detection effect is superior to the original network model and other mainstream target detection models.

关 键 词：布匹缺陷检测 YOLOv5 注意力机制 小目标缺陷 卷积操作 消融实验 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391.7[电子电信—信息与通信工程]