摘要
针对钢材表面缺陷检测中存在的精度不足以及细小缺陷极易漏检等问题,提出了一种兼顾高精度与高效率的PGL-YOLOv10n缺陷检测模型.该方法在基线YOLOv10n的基础上进行了三项核心改进:首先,引入线性可变卷积(Linear Deformable Convolution,LDConv)动态调整卷积核参数,增强网络对不同尺寸目标的自适应感知能力;其次,融入GAM注意力机制引导模型聚焦关键缺陷区域,强化局部特征表达,提升检测鲁棒性;最后,在检测头中应用独创的LPConv卷积结构,针对工业缺陷的多样性与复杂性,在保障推理速度的前提下大幅提升了特征提取效率.在NEU-DET钢材表面缺陷公开数据集上的对比实验表明,PGL-YOLOv10n模型的mAP50指标从基线的0.70显著提升至0.753.综上所述,该方法有效克服了微小瑕疵漏检的技术瓶颈,在复杂工业表面缺陷检测任务中具备显著的有效性与实际应用价值.
To address the issues of insufficient accuracy and the high susceptibility to missed detections of tiny flaws in steel surface defect detection,this paper proposes PGL-YOLOv10n,a defect detection model that achieves a balance between high accuracy and high efficiency.Based on the baseline YOLOv10n,the proposed method introduces three core improvements.First,Linear Deformable Convolution(LDConv)is introduced to dynamically adjust convolution kernel parameters,enhancing the network's adaptive perception capabilities for targets of varying sizes.Second,the GAM attention mechanism is integrated to guide the model to focus on critical defect regions,which strengthens local feature representation and improves detection robustness.Finally,an original LPConv convolution structure is applied to the detection head.Tailored to the diversity and complexity of industrial defects,it significantly boosts feature extraction efficiency while maintaining inference speed.Comparative experiments on the NEU-DET public dataset for steel surface defects demonstrate that the PGL-YOLOv10n model achieves a significant increase in the mAP50 metric,rising from the baseline of 0.70 to 0.753.In conclusion,the proposed method effectively overcomes the technical bottleneck of missed detections for tiny flaws,demonstrating substantial effectiveness and practical application value in complex industrial surface defect detection tasks.
作者
王剑
费俊桦
郭程翔
魏志杰
WANG Jian;FEI Junhua;GUO Chengxiang;WEI Zhijie(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China;Yunnan Provincial Key Laboratory of Artificial Intelligence,Kunming University of Science and Technology,Kunming 650500,China)
出处
《昆明理工大学学报(自然科学版)》
2026年第2期21-30,41,共11页
Journal of Kunming University of Science and Technology(Natural Science)
基金
云南省重大专项(202502AD080013).
