基于改进YOLOv8模型的增材制造微小气孔缺陷检测及其尺寸测量  

YOLOv8 model-based additive manufacturing micro porosity defect detection and its dimension measurement

作　　者：蔡引娣 张殿鹏 孙梓盟 王宇轩 朱祥龙 康仁科[1] CAI Yindi;ZHANG Dianpeng;SUN Zimeng;WANG Yuxuan;ZHU Xianglong;KANG Renke(College of Mechanical Engineering,Dalian University of Technology,Dalian 116024,China)

机构地区：[1]大连理工大学机械工程学院,辽宁大连116024

出　　处：《光学精密工程》2024年第21期3222-3230,共9页Optics and Precision Engineering

基　　金：国家重点研发计划资助项目(No.2022YFB4600903);国家自然科学基金联合基金集成项目(No.U23B6005);中央高校基本科研业务费资助项目(No.DUT24MS005)。

摘　　要：针对增材制造零件表面小缺陷检测中存在的准确率低和尺寸测量困难的问题,基于改进的YOLO v8提出了一种缺陷检测方法。在YOLOv8模型的检测头部引入高效通道注意力机制模块,同时采用加权交并比损失函数替换原有损失函数,减少低质量样本影响,提升检测精度。针对高分辨率图像数据集训练困难且易出现过拟合的问题,在训练阶段对包含目标缺陷的局部特征以中心为基准进行裁剪的方法生成训练集。在推理阶段,采用滑窗切分法将待测高分辨率图像裁剪成一组小图像进行预测,从而得到缺陷图像块。检测后的缺陷图像块被视为感兴趣区域,并通过计算机视觉中的边缘检测方法实现缺陷尺寸的精密测量。实验证明,改进模型的准确率达94.3%,召回率为93.3%,mAP50达到97.3%,缺陷尺寸的测量精度可达到40μm,显著提升了增材制造零件表面缺陷的检测准确率。To address challenges related to low detection accuracy and poor dimensional measurement precision of small defects on metal additive manufacturing surfaces,this study proposes a novel defect detection method based on the You Only Look Once(YOLO)v8 model.The Efficient Channel Attention(ECA)module is integrated into the detection head of the YOLOv8 framework,and the Complete Intersection Over Union(CIoU)loss function is replaced with the Wise Intersection Over Union(WIoU)loss function,effectively mitigating the impact of low-quality samples and enhancing detection performance.To overcome difficulties associated with training on high-resolution image datasets,which often lead to overfitting,local features containing target defects are cropped during the training phase to generate the training dataset.During inference,high-resolution test images are divided into smaller sub-images using a sliding window approach for defect prediction.Detected defect sub-images are marked as regions of interest,and precise defect size measurement is achieved through edge detection techniques in computer vision.Experimental results demonstrate that the improved model achieves a detection accuracy of 94.3%,a recall rate of 93.4%,and an mAP50 of 97.3%,significantly outperforming traditional methods.Furthermore,the dimensional measurement accuracy for small defects reaches 40μm,highlighting the effectiveness of the proposed approach.

关 键 词：精密测量 计算机视觉 深度学习 缺陷检测 高分辨率图片 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]