基于SDD-YOLO的轻量级带钢缺陷实时检测算法  

Lightweight strip defect real-time detection algorithm based on SDD-YOLO

作　　者：梁秀满 肖寒 LIANG Xiuman;XIAO Han(College of Electrical Engineering,North China University of Science and Technology,Tangshan 063210,China)

机构地区：[1]华北理工大学电气工程学院,河北唐山063210

出　　处：《中国测试》2025年第3期154-161,共8页China Measurement & Test

摘　　要：针对复杂工业生产环境下,热轧带钢的表面缺陷检测准确率低,网络模型参数量过大,难以部署等问题,该文提出一种基于YOLOv4模型改进的轻量级带钢缺陷实时检测算法SDD-YOLO。所提算法在特征提取部分采用GhostNet网络,压缩模型参数量;在特征融合部分,借鉴BiFPN结构改进PAN网络,采用GSConv卷积代替标准卷积,减少模型参数量和计算量,同时嵌入注意力模块CA(Coordinate Attention),增强模型特征融合能力;在预测部分采用SIOU-loss代替CIOU-loss,提高模型收敛效率,加快收敛速度;采用k-means聚类算法重新设计先验框,提高模型精度。实验结果表明,该文提出的模型相较于YOLOv4,模型参数量减少71.6%,浮点运算量降低74.6%,模型大小减小71.6%,检测精度提高3.49%,单张图片检测速度为25.9 ms。在保证准确率和检测速度的条件下,基本可以满足工业现场对缺陷的实时检测要求。The paper proposes a lightweight strip defect detection algorithm,SDD-YOLO,based on the YOLOv4 model,which uses GhostNet network in the feature extraction part to compress the number of model parameters.In the feature fusion part,the PAN network is improved with the BiFPN structure.And GSConv convolution is used instead of standard convolution to reduce the number of model parameters and computational effort.At the same time,the attention module CA is embedded to enhance the model's feature fusion ability.SIOU-loss is used instead of CIOU-loss in the prediction part to improve the model convergence efficiency and accelerate the convergence speed.The k-means clustering algorithm is used to redesign the prior frame to improve the model accuracy.The experimental results show that the model proposed in this paper,compared with YOLOv4,reduces the amount of model parameters by 71.6%,the amount of floating-point operations by 74.6%,the model size by 71.6%,the detection accuracy by 3.49%,and the detection speed of a single image by 25.9 ms.Under the condition of ensuring the accuracy and detection speed,it can basically meet the requirements of real-time detection of defects in industrial sites.

关 键 词：带钢表面缺陷 目标检测 轻量级网络 YOLOv4 

分 类 号：TB9[一般工业技术—计量学] TG335.56[机械工程—测试计量技术及仪器] TP183[金属学及工艺—金属压力加工]