基于轻量化YOLOv8s算法的金属表面缺陷检测研究  

Metal surface defect detection based on lightweight YOLOv8s algorithm

作　　者：王少杰 朱皓然[1,2] WANG Shaojie;ZHU Haoran(Hunan Provincial Key Laboratory of Grids Operation and Control on Multi-Power Sources Area,Shaoyang 422000,China;School of Electrical Engineering,Shaoyang University,Shaoyang 422000,China)

机构地区：[1]多电源地区电网运行与控制湖南省重点实验室,湖南邵阳422000 [2]邵阳学院电气工程学院,湖南邵阳422000

出　　处：《邵阳学院学报(自然科学版)》2025年第2期11-18,共8页Journal of Shaoyang University:Natural Science Edition

基　　金：湖南省科技厅重点研发计划项目(2024JK2040);邵阳学院研究生创新项目(CX2023SY051)。

摘　　要：针对金属原材料表面的缺陷检测问题,设计了一种基于轻量化YOLOv8s的改进模型。首先,对YOLOv8s模型进行稀疏化训练,并对归一化(batch normalization,BN)层进行通道剪枝,降低模型的参数量。同时,因为对模型剪枝后,损失了一定的精度,所以使用知识蒸馏技术辅助模型并且微调,将模型的精度拉到剪枝之前的精度左右。将损失函数替换为一种基于归一化Wasserstein距离的新度量(normalized Wasserstein distance,NWD),有效提升模型检测小目标的性能。再加入极化自注意力机制(polarized self-attention,PSA),可以有效地完成特征融合。实验数据表明,改进后的YOLOv8s模型在GC10-DET数据集上的平均精度相比原来的网络模型提高了2.3%,模型参数量下降了23.3%,召回率提高了4.6%,能够很好地完成对金属表面缺陷进行检测的需求。In order to solve the problems in defect detection on the surface of metal materials,An improved model based on lightweight YOLOv8s was designed.Firstly,sparsity training was carried out for the YOLOv8s model,and the batch normalization(BN)layer was pruned to reduce the number of parameters of the model.The knowledge distillation technology was employed for fine-tuning to mitigate the accuracy loss due to pruning,thereby restoring accuracy close to pre-pruning levels.The loss function was replaced by a new metric based on the normalized Wasserstein distance(NWD),which effectively improved the model’s performance in detecting small targets.Then,a polarized self-attention(PSA)mechanism was introduced to complete feature fusion effectively.Experimental results indicated that the improved YOLOv8s model achieved a 2.3%increase in mean average precision,a 23.3%reduction in parameters,and a 4.6%improvement in recall rate on the GC10-DET dataset compared to the original model,demonstrating its efficacy for metal surface defect detection.

关 键 词：金属缺陷检测 YOLOv8s 轻量化网络 极化自注意力机制 损失函数 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]