一种基于改进YOLOv8的带状合金功能材料缺陷检测方法  

A Defect Detection Method for Strip Alloy Functional Materials Based on Improved YOLOv8

作　　者：杨威 杨俊[1] 许聪源 夏亚金 YANG Wei;YANG Jun;XU Congyuan;XIA Yajin(School of Information Science and Engineering,Jiaxing University,Jiaxing,Zhejiang 314001,China;School of Computer Science and Technology,Zhejiang Sci-Tech University,Hangzhou,Zhejiang 310018,China;Haiyan Zhongda METAL Electronic Material Co.LTD,Jiaxing,Zhejiang 314300,China)

机构地区：[1]嘉兴大学信息科学与工程学院,浙江嘉兴314001 [2]浙江理工大学计算机科学与技术学院,浙江杭州310018 [3]海盐中达金属电子材料有限公司,浙江嘉兴314300

出　　处：《计量学报》2025年第3期329-339,共11页Acta Metrologica Sinica

基　　金：国家自然科学基金(62302197);浙江省自然科学基金(LQ23F020006);嘉兴市科技计划基金(2024AD10045)。

摘　　要：针对带状合金功能材料缺陷检测中存在的漏检、误检和检测速度慢等问题,提出一种基于改进YOLOv8的带状合金功能材料缺陷检测算法。为充分融合模型骨干网络提取的多尺度特征,首先,设计多尺度特征编码器(MFE)模块,并在颈部构建多尺度特征聚集扩散(MFAD)结构,利用独特的扩散机制使具有丰富上下文信息的特征扩散到各个尺度;然后,在模型头部设计一种共享参数的任务动态对齐检测头(TDADH),通过卷积参数共享与任务对齐机制,降低模型复杂度的同时提高模型的检测精度;最后,设计感知注意力空间金字塔池化(PASPP)模块,利用注意力机制的显式动态选择机制增强模型特征表达能力。实验结果表明:该方法在合金功能材料数据集上实现了90.1%的均值平均精度P_(mAP50),参数量为2.543×10^(6),检测速度为232帧/s,优于主流的深度检测算法,并在GC10-DET和PASCAL VOC2012数据集上获得最优性能,具备较好的泛化性。In order to solve the problems of missed detection,false detection,and slow detection speed in the defect detection of strip alloy functional materials,a defect detection algorithm for strip alloy functional materials based on improved YOLOv8 is proposed.In order to fully integrate the multi-scale features extracted by the model backbone network,a multi-scale feature encoder(MFE)module is first designed,and a multiscal feature affregation-diffusion(MFAD)structure is constructed at the neck.The unique diffusion mechanism is used to diffuse features with rich contextual information to all scales.Then,a shared parameter task dynamic alignment detection head(TDADH)is designed at the head of the model.Through convolution parameter sharing and task alignment mechanisms,the model complexity is reduced while the detection accuracy is improved.Finally,a perceptual attention spatial pyramid pooling(PASPP)module is designed to enhance the feature expression ability of the model using the explicit dynamic selection mechanism of attention mechanism.Experimental results indicate that the method proposed attains a mean average precision(PmAP50)of 90.1%on the alloy functional material dataset.It boasts a parameter count of 2.543×10^(6) and a detection speed of 232 FPS(Frames Per Second),outperforming leading deep detection algorithms.Moreover,it achieves top performance on the GC10-DET and PASCAL VOC2012 datasets,demonstrating strong generalizability ability.

关 键 词：机器视觉检测 表面缺陷检测 带状合金功能材料 多尺度融合 解耦检测头 注意力机制 YOLOv8 

分 类 号：TB96[机械工程—光学工程] TB973[一般工业技术—计量学]