文献类型：期刊文献

中文题名：基于GSL-YOLO的钢材表面缺陷检测算法

英文题名：Steel Surface Defect Detection Algorithm Based on GSL-YOLO

作者：王泽恺[1];姜庆超[1];颜学峰[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237

年份：2025

卷号：51

期号：5

起止页码：657

中文期刊名：华东理工大学学报(自然科学版)

外文期刊名：Journal of East China University of Science and Technology

收录：;北大核心:【北大核心2023】；

基金：装备预研教育部联合基金(8091B032230);国家自然科学基金优秀青年基金(62322309)。

语种：中文

中文关键词：缺陷检测;YOLOv8;Ghost Bottleneck;DynamicConv;注意力机制;轻量级检测头

外文关键词：defect detection;YOLOv8;Ghost Bottleneck;DynamicConv;attention mechanism;lightweight detection head

摘要：钢材在现代工业生产中因加工和环境等因素易产生表面缺陷,影响其质量和使用寿命。针对现有检测方法存在的精度不足与计算复杂度较高的问题,本文提出了一种基于YOLOv8的改进算法GSL-YOLO,针对特征冗余和特征表达能力不足的问题,在C2f模块中引入Ghost Bottleneck替代标准Darknet Bottleneck,减少冗余计算并提高特征提取效率,采用DynamicConv替换标准CBS模块,以增强对复杂场景的适应能力。为进一步提升模型对细小缺陷的检测精度,在颈部网络中引入SENetV1注意力机制,通过自适应调整通道权重,加强对关键特征的关注。此外,针对检测头参数量大、计算复杂度高的问题,设计轻量级共享卷积检测头LSDECD,有效降低模型参数量和计算开销。实验结果表明,GSL-YOLO在NEU-DET钢材表面缺陷数据集上的mAP50较YOLOv8提升3%,同时参数量减少33%,计算成本降低37%。
Steel is prone to surface defects due to processing and environmental factors in modern industrial production,which adversely affect its overall quality and service life.To overcome the shortcomings of existing detection methods,such as,inadequate accuracy and high computational complexity,this paper proposes an improved YOLOv8-based algorithm,GSL-YOLO.To address the issues of feature redundancy and insufficient feature expression ability,GhostBottleneck is introduced in the C2f module to replace the standard DarknetBottleneck,thereby reducing redundant computation and enhancing feature extraction efficiency.DynamicConv is used to replace the standard CBS module to enhance adaptability to complex scenarios.To further improve the detection accuracy for minute defects,the SENetV1 attention mechanism is integrated into the neck network,which enhances the focus on key features by adaptively adjusting channel weights.In addition,a lightweight shared convolution detection head named LSDECD is designed to address the issues of large number of detection head parameters and high computational complexity,effectively reducing the number of model parameters and the cost of computation.The experimental results demonstrate that,on the NEU-DET steel surface defect dataset,GSL-YOLO achieves a 3%increase in mAP50 compared with YOLOv8,while he parameter quantity is reduced by 33%and the computational cost is reduced by 37%simultaneously.The proposed method achieves good lightweight design while improving the accuracy of model detection,meeting the application requirements of industrial real-time detection.