基于YOLOv7-NBC的钢材表面缺陷检测算法研究  

Research on Steel Surface Defect Detection Algorithm Based on YOLOv7-NBC

作　　者：李淇 石艳 林峰 郝琪 LI Qi;SHI Yan;LIN Feng;HAO Qi(School of Mechanical Engineering,Sichuan University of Science&Engineering,Yibin 644000,China)

机构地区：[1]四川轻化工大学机械工程学院,四川宜宾644000

出　　处：《四川轻化工大学学报(自然科学版)》2025年第1期47-56,共10页Journal of Sichuan University of Science & Engineering(Natural Science Edition)

基　　金：四川省科技计划重点研发项目(2022YFG0068);四川省大学生创新创业训练项目(CX2023074)。

摘　　要：针对钢材表面小尺寸缺陷与复杂背景之间的区别度太低导致检测效果不佳的问题,提出了一种基于YOLOv7-NBC的钢材表面缺陷检测算法,NBC分别代表引入的NWD度量标准、动态稀疏注意力模块(BiFormer)和级联融合网络结构(Cascade Fusion Network,CFNet)。主要改进如下:在YOLOv7算法的骨干网络第24层引入动态稀疏注意力模块,提高算法的特征学习能力;通过寻求IoU度量标准与NWD度量标准的最优比值,获得更好的损失以降低钢材表面缺陷位置的偏差敏感性,提高算法对缺陷的检测性能;在骨干网络处引入级联融合网络结构,减少算法参数量。并进一步将改进后的YOLOv7-NBC算法应用于优化后的NEU-DET数据集上做消融与对比实验。实验结果表明,与YOLOv7相比,YOLOv7-NBC算法的检测精度有明显提升,mAP达到了85.4%,提升了4.6%;YOLOv7-NBC算法的计算量降低了52.1%,FPS达到70,提高了工业检测效率。YOLOv7-NBC算法具有更高的检测精度,泛化能力更强,错误和漏检率更低。To solve the problem that the difference between small size defects on steel surface and complex background is too low,leading to poor detection effect,a steel surface defect detection algorithm based on YOLOv7-NBC is proposed,with NBC representing the introduced NWD metric,BiFormer and Cascade Fusion Network(CFNet)respectively.The dynamic sparse attention module is introduced in the 24th layer of the backbone network of YOLOv7 algorithm to improve the feature learning ability of the algorithm.By seeking the optimal ratio of IoU metric to NWD metric,better loss is obtained to reduce the bias sensitivity of steel surface defect location and improve the detection performance of the algorithm.The cascade fusion network structure is introduced into the backbone network to reduce the number of algorithm parameters.The improved YOLOv7-NBC algorithm is applied to the optimized NEU-DET dataset for ablation and comparison experiments.The experimental results show that the detection accuracy of the YOLOv7-NBC algorithm is significantly improved,with mAP reaching 85.4%,an increase of 4.6%.The calculation amount of YOLOv7-NBC algorithm is reduced by 52.1%,the FPS reaches 70,and the industrial detection efficiency is improved.YOLOv7-NBC algorithm has higher detection accuracy,stronger generalization ability,and lower error and missed detection rate.

关 键 词：复杂背景 小尺寸缺陷 缺陷检测 YOLOv7 动态稀疏注意力模块 级联融合网络结构 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]