基于改进YOLOv8的织物疵点检测算法  

Fabric Defect Detection Algorithm Based on Improved YOLOv8

作　　者：罗维平 张哲 Luo Weiping;Zhang Zhe(College of Mechanical Engineering and Automation,Wuhan Textile University,Wuhan City,Hubei Provincce 430200)

机构地区：[1]武汉纺织大学机械工程与自动化学院,湖北武汉430200

出　　处：《黄河科技学院学报》2025年第2期23-30,共8页Journal of Huanghe S&T College

摘　　要：针对织物疵点检测中疵点形态各异,检测结果容易存在漏检或误检等问题,提出一种改进YOLOv8的织物疵点检测算法。首先将主干网络替换为GhostNet,该方法在减少对计算资源要求的同时,在保证网络结构质量的前提下,减轻了网络结构的重量。另外,该模型还利用了一系列有效的线性运算,减少了模型中的参数,获得更多的特征图谱,减少了算法的计算量,大大提高了模型的性能。其次引入CA注意力机制(coordinate attention),有效增强了网络的特征提取能力。最后引入BiFPN金字塔替换head层中的concat连接,将语义信息传递到不同的特征尺度上,从而增强特征融合。实验结果表明改进的算法在检测8类织物疵点时,mAP@0.5达到91.9%,mAP@0.5:0.95达到48.8%,相比原始的YOLOv8n算法分别提高了2.2%和1.1%,具备更高的检测精度,具有更少的漏检或误检情况。A fabric defect detection algorithm based on improved YOLOv8 is proposed to address the issue of various defect shapes and the potential for missed or false detections in fabric defect detection results.Firstly,the backbone network was replaced with GhostNet,which not only reduces the demand for computational resources but also achieves a lightweight network structure without compromising model performance.In addition,the Ghost module generates more feature maps with fewer parameters through a series of efficient linear operations,significantly reducing the number of parameters and computational complexity,making it more advantageous compared to traditional convolutional neural networks.Secondly,the introduction of CA attention mechanism effectively enhances the feature extraction capability of the network.Finally,BiFPN pyramids are introduced to replace the concat connections in the head layer,conveying semantic information to different feature scales and enhancing feature fusion.The experimental results show that the improved algorithm is effective in detecting 8 types of fabric defects,m AP@0.5 Reaching 91.9%,m AP@0.50.95 reaches 48.8%,which is 2.2%and 1.1%higher than the original YOLOv8n algorithm,respectively.It has higher detection accuracy and fewer missed or false detections.

关 键 词：疵点检测 YOLOv8算法 GhostNet 坐标注意力 BiFPN 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]