多尺度过滤融合印刷电路板缺陷检测方法  

Multi-scale filtering fusion printed circuit board defect detection method

作　　者：余泽寰 陈锐 吕宇 YU Zehuan;CHEN Rui;LYU Yu(School of Mechanical and Electronic Engineering,East China University of Technology,Nanchang 330013,China)

机构地区：[1]东华理工大学机械与电子工程学院,南昌330013

出　　处：《激光杂志》2025年第4期79-87,共9页Laser Journal

基　　金：国家自然科学基金(No.12365026);江西省重点研发计划重点项目(No.20232BBE50013);抚州市揭榜挂帅项目(No.XMBH00016)。

摘　　要：针对当前印刷电路板检测算法性能要求过高,缺陷较难识别的问题,提出一种多尺度过滤融合印刷电路板缺陷检测方法。方法设计了加权过滤特征金字塔网络(WF-FPN),通过简化融合路径,优化不同层次间信息传递效率,显著降低模型复杂度并提升检测性能。同时,在主干网络中加入可变形注意力,通过生成偏移子网络增强对不规则缺陷的特征提取能力。并采用完整瓦瑟斯坦距离作为损失函数加快训练收敛速度、提高精确度。最后,使用轻量化检测头降低模型性能开销。实验结果表明,改进后的模型权重文件大小仅为3.62 MB对比基线模型下降39.4%,同时计算量下降35.8%,精度提升2.1%。与同类算法相比能更好地满足模型在嵌入式、移动式检测设备中部署的需要。Aiming at the current printed circuit board detection algorithm performance requirements are too high,defects are more difficult to recognize the problem,a multi-scale filtering fusion printed circuit board defect detection method is proposed.The method designs a weight-filtering feature pyramid network(WF-FPN),which significantly reduces the model complexity and improves the detection performance by simplifying the fusion paths and optimizing the information transfer efficiency between different levels.At the same time,deformable attention is added to the backbone to enhance the feature extraction capability for irregular defects by generating offset subnetworks.The complete Wasserstein distance is used as the loss function to accelerate the training convergence speed and improve the accuracy.Finally,a lightweight detection head is used to reduce the model performance overhead.Experimental results show that the size of the improved model weight file is only 3.62 MB which is 39.4%lower than the baseline model,while the computation amount is 35.8%lower and the accuracy is 2.1%higher.Compared with similar algorithms,it can better meet the needs of the model deployed in embedded and mobile detection devices.

关 键 词：印刷电路板 缺陷检测 YOLOv8 特征融合金字塔 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] TN41[自动化与计算机技术—计算机科学与技术]