基于改进HRNet的钢材缺陷像素级检测算法  

Pixel-level Detection Algorithm for Steel Defects Based on Improved HRNet

作　　者：孙英伟 张岩[1] SUN Yingwei;ZHANG Yan(College of Electromechanical Engineering,Qingdao University of Science and Technology,Qingdao 266061)

机构地区：[1]青岛科技大学机电工程学院,青岛266061

出　　处：《计算机与数字工程》2025年第3期840-844,920,共6页Computer & Digital Engineering

摘　　要：论文提出了一种基于HRNet改进的高分辨率感知网络,用于像素级的检测钢材生产过程中产生的缺陷。该算法既可以定位缺陷的位置,也能够表征缺陷的几何形状。基于HRNet,论文设计了一种多尺度注意力感知模块(MAAM),用于增强每个阶段之间特征的信息交互,通过通道注意力和空间注意力感知特征融合后的重要信息。另外,论文提出了一种混合损失函数,用于监督预测结果与真实标签的差距,提高了对钢材缺陷的检测准确率。经过实验验证,该算法能够应对各种类型的钢材缺陷,并且具有较高的检测准确率。This paper proposes an improved high-resolution perception network based on HRNet for pixel-level detection of defects in steel production.The algorithm can not only locate the location of the defect,but also characterize the geometry of the defect.Based on HRNet,this paper designs a multi-scale attention-aware module(MAAM)to enhance the information interaction of features between each stage,and perceive important information after feature fusion through channel attention and spatial attention.In addition,this paper proposes a hybrid loss function to supervise the gap between the predicted results and the true labels,which improves the detection accuracy of steel defects.After experimental verification,the algorithm can deal with the defects of various types of steel,and has a high detection accuracy.

关 键 词：缺陷检测 钢材缺陷 注意力机制 高分辨率网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程] TG142[一般工业技术—材料科学与工程]