用于缺陷检测的YOLOv8轻量化设计方法  

YOLOv8 lightweight design approach for defect detection

作　　者：艾峰 邓耀华[1] Ai Feng;Deng Yaohua(College of Mechanical and Electrical Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学机电工程学院,广州510006

出　　处：《电子测量技术》2025年第4期181-190,共10页Electronic Measurement Technology

基　　金：广东省省级科技计划项目(2023A0505050151)资助。

摘　　要：在大规模制造的端侧产线工业质检应用中,由于算力、成本和功耗等因素的限制,将深度学习模型裁剪并部署到小型算力的边缘设备上变得尤为重要。针对铝型材复杂缺陷检测这一应用场景,基于YOLOv8设计了缺陷检测模型。首先,通过轻量化结构设计,结合局部自注意力机制提升细微缺陷提取能力;采用空间通道下采样替代传统下采样卷积;并提出结合混合局部通道注意力机制的C2f-M模块。然后,基于双向特征金字塔网络设计了SC-BiFPN颈部网络,增强了多尺度特征融合能力。接着,设计任务动态对齐的特征检测头TDADH,充分利用多层次特征,实现更精准的目标定位与分类;采用MPDIoU损失函数增强边界框回归能力。最后,通过Taylor方法对YOLOv8进行裁剪,显著减少模型参数量和计算成本。实验结果表明,轻量化YOLOv8模型在铝材表面缺陷数据集上的参数量降低至原模型的36.7%,计算量减少40%,模型体积缩小62%;同时,检测精确度、召回率及mAP@50-95分别提升0.3%、1.1%、4.8%。该方法有效解决了端侧部署中的计算复杂度与检测性能平衡问题,为小型算力硬件上的高效缺陷检测提供了可行方案。In end-side production line industrial quality inspection applications for large-scale manufacturing,tailoring and deploying deep learning models to edge devices with small arithmetic power becomes particularly important due to the limitations of arithmetic power,cost and power consumption.Based on the application scenario of complex defect detection of aluminum profile,the defect detection model is designed based on YOLOv8.First of all,through the lightweight structure design,combined with the partial self-attention mechanism to improve the ability of subtle defect extraction;the use of spatial channel downsample instead of the traditional downsampling convolution;and proposed a combination of mixed local channel attention mechanism of the C2f-M module.Then,SC-BiFPN neck network is designed based on bidirectional feature pyramid network,which enhances the multi-scale feature fusion capability.Then,the task dynamic align detection head is designed to make full use of multilevel features for more accurate target localisation and classification;the MPDIoU loss function is used to enhance the bounding box regression capability.Finally,YOLOv8 is trimmed by Taylor's method to significantly reduce the number of model parameters and computational cost.The experimental results show that the lightweight YOLOv8 model reduces the number of parameters to 36.7%of the original model on the aluminium surface defects dataset,reduces the computational effort by 40%,and reduces the model volume by 62%;at the same time,the detection accuracy,the recall rate and the mAP@50-95 are improved by 0.3%,1.1%,and 4.8%,respectively.The method effectively solves the problem of balancing computational complexity and detection performance in end-side deployment,and provides a feasible solution for efficient defect detection on small arithmetic hardware.

关 键 词：缺陷检测 双向特征金字塔 损失函数 任务动态对齐 剪枝 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN40[自动化与计算机技术—计算机科学与技术]