摘要
目的针对现有带钢表面缺陷检测精度较低、存在漏检和误检等问题,提出了一种改进YOLOv8的缺陷检测算法YOLOv8-SSDW。方法该算法以YOLOv8n为基准模型,在骨干网络结构中引入SKNet(Selective Kernel Networks)注意力模块,加强骨干网络的特征提取能力和自适应能力,使网络在特征提取过程中更关注缺陷目标;同时,在YOLOv8的颈部使用Slim-Neck结构,减少模型的参数量和计算量;为进一步提升网络的特征提取能力,提出一种融合可变形卷积,强化对缺陷目标的特征学习;考虑缺陷样本质量不平衡问题,使用WIoU(wise intersection over union)损失函数,其梯度增益分配策略使问题得到有效改善,并且提高模型收敛速度和回归精度。结果改进后的模型在带钢数据集上进行实验,结果表明:改进后的模型的平均精度达到85.5%,相比基准模型提高了2.7%。结论通过大量实验可以证明改进网络的有效性,改善了带钢表面缺陷检测精度较低的问题,减少了漏检和误检的情况,同时满足实时性要求;相较于目前主流模型,该改进算法在检测精度具有一定优势,对后续研究用于实际检测具有参考价值。
Objective In response to the issues of low detection accuracy,missed detections,and false alarms in existing steel surface defect detection methods,an improved defect detection algorithm,YOLOv8-SSDW,based on YOLOv8,was proposed.Methods This algorithm took YOLOv8n as the benchmark model and introduced the SKNet(Selective Kernel Networks)attention module into the backbone network structure to enhance the feature extraction and adaptability of the backbone network,allowing the network to pay more attention to defect targets during the feature extraction process.At the same time,the Slim-Neck structure was used in the neck of YOLOv8 to reduce the number of model parameters and computational load.To further enhance the network’s feature extraction capability,a deformable convolution fusion method was proposed to strengthen the feature learning for defect targets.Considering the imbalance in defect sample quality,the WIoU(wise intersection over union)loss function was used,which effectively addressed the issue through its gradient gain allocation strategy,enhancing model convergence speed and regression accuracy.Results Experiments on the steel dataset showed that the average accuracy of the improved model reached 85.5%,which was an increase of 2.7%over the benchmark model.Conclusion Extensive experiments demonstrate the effectiveness of the improved network,which resolves the issue of low accuracy in steel strip surface defect detection,reduces missed and false detections,and meets real-time requirements.Compared with current mainstream models,the proposed model has certain advantages in detection accuracy and offers a valuable reference for practical detection in future research.
作者
戴林华
黎远松
石睿
DAI Linhua;LI Yuansong;SHI Rui(School of Computer Science and Engineering,Sichuan University of Science&Engineering,Sichuan Yibin 643002,China)
出处
《重庆工商大学学报(自然科学版)》
2025年第4期44-52,共9页
Journal of Chongqing Technology and Business University(Natural Science Edition)
基金
国家自然科学基金资助项目(42074218)。
作者简介
戴林华(1999-),男,江西九江人,硕士,从事目标检测研究.Email:3228958982@qq.com;通信作者:黎远松(1970-),男,教授,硕士生导师,从事地图学与地理信息系统、在线实时测量与无损检测、机器学习等研究.Email:yuansongli@suse.edu.cn.
