微型元器件引脚测距的深度学习方法研究
2025-08-15项朝辉钱月钟邹佳铭
中图分类号:TP391.41 文献标识码:A 文章编号:1672-0105(2025)02-0047-05
Research on Deep Learning Methods for Micro-component Pin Distance Measurement
XIANG Chaohui, QIAN Yuezhong, ZOU Jiaming (ZhejiangIndustryTradeVocationalCollege,Wenzhou325o35,China)
Abstract:Inindustrial visualquality inspection,measuring the distancebetween pinsof microelectroniccomponents posesa significantchalenge.TraditionalmethodsrelyheavilyonOpenCVforextensive imageprocesingoperations,whichrequieegiees torepeatedlyadjustparameters,leading toloweiciencyandinsuficientauracyIntherealmofdeepleaing,theseisuscanbe addressedthrough distanceregresionandobject detectiontechniques.MethodOneivolvesusingtheResNet18 network to extract image features and employing fullyconnectedlayers fordistanceregresion toultimately measure pin distances.Method Two trains the YOLOv8n modelto detect pin bounding boxes and then calculates the distances between pins basedon these detections. Aditionallyduetoteghomogeneityofimagescolectedfromproductionlinswherepinpositionsarelargelysimlaodels oftensuffer fromsevereoverfiting.Tomitigatethis,wedesignedasimulated pindistancemeasurementmethodtogeneratediverse data,therebyenhancingtherobustnessofthemodels.Experimentalresultsdemonstratethatthedep learing-baseddistance measurement methods offer a significant improvement in acuracy compared to traditional OpenCV-based approaches.
Keywords: industrial quality inspection; deep learning; object detection; pin distance measurement
0引言
微型电子元器件是电子系统中的关键组件,其质量影响整个系统的性能。随着工业生产的快速发展,微型电子元器件的生产效率大幅提升,但表面缺陷(如填充不足、裂痕、缺焊片、划痕等)和尺寸缺陷(如引脚间距过宽/窄、焊片长度过长/短等)等问题依然存在。鉴于这些元件价格低廉且体积小巧,对单个元器件的高效精确检测成为必要,然而人工审核难以满足速度与精度的要求。近年来,随着人工智能技术的进步,基于机器视觉的检测方法逐渐成为微型电子元器件质检的主要手段。
在工业视觉检测中,Zheng等结合OpenCV的OTSU自动阀值处理和Canny边缘检测,提取了胶囊外部轮廓,并通过形态特征分析来识别内部缺陷。He等2采用不对称侵蚀扩张技术进行背景区域分割,提取不受阴影干扰的特征区域,并基于菱形点矫正(RSPC)算法和点重排技术,有效检测了电力网络中的缺陷。WeiFan等聚焦车辆表面缺陷检测,将快速二维傅里叶变换(2DFFT)应用于漫射照明摄像系统中,实验表明算法效果良好。Keyur及其团队4在齿轮、硬币和连接器的分类中,提出了融合支持向量机以及神经网络的分类方法,使准确率达到 95% 。DavidFries及其团队针对飞行过程中飞行员遭遇不明原因生理事件(UPEs)所带来的安全隐患,开发了一套快速响应的机器视觉解决方案,该系统集成了先进的机器学习算法和认知神经网络技术,能够实时分析来自多种传感器的数据流。……
