改进YOLOv8算法的钢材表面缺陷检测  

Detection of Steel Surface Defect Using Improved YOLOv8 Algorithm

作　　者：王孟宇 刘志强[1,2] WANG Mengyu;LIU Zhiqiang(College of Automotive and Mechanical Engineering,Changsha University of Science&Technology,Changsha 410114,China;College of Excellent Engineers,Changsha University of Science&Technology,Changsha 410114,China)

机构地区：[1]长沙理工大学汽车与机械工程学院,长沙410114 [2]长沙理工大学卓越工程师学院,长沙410114

出　　处：《机械科学与技术》2025年第1期19-29,共11页Mechanical Science and Technology for Aerospace Engineering

摘　　要：针对钢材表面缺陷尺度不一,现有检测算法多尺度特征处理能力较差、精度有待提高的问题,提出了一种面向钢材表面缺陷检测的改进型算法,命名为ADP-YOLOv8。首先,提出了一种自适应权重下采样(Adaptive weight downsampling)模块,其通过加权组合不同的下采样特征图,增强了模型对缺陷信息的关注;然后,通过改进特征提取网络中的C2F模块,加强从网络高层的可扩展感受野中提取特征;最后,引入可编程梯度信息(PGI)模块,通过其多级辅助信息组件逐步整合不同尺度的特征,有效提高了模型对不同尺度缺陷敏感性。所提出方法的平均精度为79.3%,相比基准模型提高了3.5%;检测速度为163.2frame/s。相比其他主流的目标检测算法,改进后的检测器在性能上更具优势,展示出了在检测精度、速度和模型体积方面的良好平衡。A modified algorithm for steel surface defect detection,named ADP-YOLOv8,is proposed to address the issues of uneven scale of steel surface defects,poor multi-scale feature processing ability of existing detection algorithms,and the need to improve accuracy.Firstly,an adaptive weighted downsampling(ADSConv)module is proposed,which enhances the detector′s adaptability to different types of defects by weighting and combining different downsampling feature maps.Then,by improving the C2F module in the feature extraction network,the extraction of features from the scalable receptive field at the higher level of the network is strengthened.Finally,the introduction of the programmable gradient information(PGI)module gradually integrates features of different scales through its multi-level auxiliary information components.The average accuracy of the present method is 79.3%,which is 3.5%higher than the benchmark model.The detection speed is 163.2 frame/s.Comparing with the other mainstream object detection algorithms,the improved detector has more advantages in performance,demonstrating a good balance in detection accuracy,speed and model volume.

关 键 词：表面缺陷检测 自适应权重 感受野 可编程梯度信息 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]