基于改进YOLOv8n的织物疵点检测  

Fabric defect detection based on improved YOLOv8n

作　　者：李耀 徐红伟[1] 柯海森[1] 郭殿鹏 李孝禄[1] LI Yao;XU Hongwei;KE Haisen;GUO Dianpeng;LI Xiaolu(China Jiliang University,Hangzhou,310018,China)

机构地区：[1]中国计量大学,浙江杭州310018

出　　处：《棉纺织技术》2024年第10期11-18,共8页Cotton Textile Technology

基　　金：浙江省自然科学基金重点项目(LZ23E060002)。

摘　　要：针对纺织行业中织物疵点大小不一、织物表面图案复杂等问题,提出一种基于改进YOLOv8n的织物疵点检测算法。首先,在主干网络中用CAA模块替换C2f中的Bottleneck模块,应用两个深度条形卷积作为标准大核深度卷积的近似值来减少计算量,并捕获疵点图像多尺度特征;其次,添加LSKA注意力机制,在特征提取中减少织物表面复杂图案的干扰,提升对小目标的检测精度;最后,使用MPDIoU损失函数弥补原始损失函数的局限性,提高模型训练效率。试验结果表明:改进的YOLOv8n模型mAP值达到90.2%,相比于原始YOLOv8n模型提升了5.9个百分点,同时检测速度达到73帧/s。将改进的模型部署至试验平台进行测试,mAP值和检测速度分别为87.4%和65帧/s,可满足纺织企业对织物疵点检测准确性和实时性的需求。Aiming at the problems of different fabric defect sizes and complex fabric surface patterns in the textile industry,a fabric defect detection algorithm based on improved YOLOv8n was proposed.Firstly,the Bottleneck module in C2f was replaced by CAA module in the backbone network,and two deep bar convolutions were used as the approximate value of the standard large kernel deep convolution to reduce the computational load and capture the multi-scale features of defective images.Secondly,LSKA attention mechanism was added to reduce the interference of complex patterns on fabric surface and improve the detection accuracy of small targets.Finally,MPDIoU loss function was used to make up the limitation of the original loss function and improve the efficiency of model training.The experimental results showed that the mAP value of the improved YOLOv8n model was reached 90.2%,which was 5.9 percentage points higher than that of the original YOLOv8n model,and the detection speed was reached 73 frames/s.The improved model was deployed to the test platform,and the mAP value and detection speed were 87.4%and 65 frames/s,respectively,which could meet the needs of textile enterprises for the accuracy and real-time detection of fabric defects.

关 键 词：织物疵点 YOLOv8n 深度卷积 注意力机制 损失函数 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TS101.8[自动化与计算机技术—计算机科学与技术]