摘要
钢材表面缺陷检测是保障制造业产品质量的关键环节,人工目视与基础光学检测方法存在效率低、漏检率高等问题,且现有数据集样本有限制约模型泛化能力。为此,本文提出一种基于LS-DCGAN数据增强与改进YOLOv8的轻量化钢材表面缺陷检测方法。首先针对NEU-DET数据集样本多样性不足的问题,采用LS-DCGAN生成对抗网络进行数据增强,有效补充缺陷样本的形态特征与分布特性;其次对YOLOv8模型进行三重优化提出SPH-YOLO检测算法:重构C2f模块结构增强特征提取能力,嵌入注意力机制提升缺陷区域聚焦度,设计多级特征融合金字塔实现跨尺度信息交互;最后在增强后的NEU-DET与GC10-DET数据集上进行验证,实验表明改进模型在mAP@50%指标上提升3%,参数量减少28.5%,计算量降低12.3%,且改进方法具有泛化能力,检测有效地实现了检测模型轻量化和检测性能的平衡。
Detection of surface defects in steel materials is a key link in ensuring the quality of products in the manufacturing industry.Manual visual inspection and basic optical detection methods suffer from low efficiency and high miss detection rates,and the limited samples of existing datasets restrict the model′s generalization ability.Therefore,this paper proposes a lightweight steel surface defect detection method that integrates LS-DCGAN data augmentation with an improved YOLOv8 model.Firstly,to address the issue of insufficient sample diversity in the NEU-DET dataset,we use an LS-DCGAN generative adversarial network for data augmentation,effectively supplementing the morphological features and distribution characteristics of defect samples.Secondly,we conduct triple optimization on the YOLOv8 model to propose the SPH-YOLO detection algorithm:reconstructing the C2f module structure to enhance feature extraction capabilities,embedding an attention mechanism to improve focus on defect areas,and designing a multi-level feature fusion pyramid for cross-scale information interaction.Finally,we validate the improved model on the enhanced NEU-DET and GC10-DET datasets.Experimental results show that the improved model achieves a 3.2%increase in mAP@50%,a 28.5%reduction in parameter count,and a 12.3%decrease in computational load.Furthermore,the improvement method exhibits strong generalization ability,effectively balancing the lightweight nature of the detection model and its detection performance.
作者
胥姜苗
曹爽
管海燕
Xu Jiangmiao;Cao Shuang;Guan Haiyan(School of Remote Sensing and Geomatics Engineering Institute of Optics and Fine Mechanics,Nanjing University of Information Science and Technology,Nanjing 210044,China)
出处
《电子测量技术》
2025年第24期138-147,共10页
Electronic Measurement Technology
基金
国家自然科学基金面上项目(41971414)资助。
