摘要
针对现有印刷电路板组装(printed circuit board assembly,PCBA)元件表面的缺陷检测算法存在的检测精度不足、模型复杂度高等问题,提出一种基于目标识别与特征重建的两阶段高精密光学PCBA元件表面缺陷检测方法。该方法包括元件识别和缺陷检测两阶段:首先,元件识别阶段采用目标检测算法定位PCBA的元件区域,并通过预处理与空间重叠裁剪减少全局背景噪声干扰,保证后续输入特征的语义一致性,从而提升缺陷检测的准确性与推理效率;其次,缺陷检测阶段采用基于特征重建的异常检测框架,改进预训练特征提取模型以适应随机形状和多尺度缺陷,同时可在通道和空间维度上提取更丰富的特征信息。此外,针对工业检测高处理速度的需求,提出一种结合深度可分离卷积的轻量级编码器,以降低模型参数,增强模型泛化能力。在自制PCBA数据集和MvTec公开数据集上的实验结果表明,相较于现有先进的无监督缺陷检测算法,本文方法在图像级与像素级异常检测性能上均取得最优结果,并显著降低模型复杂度,满足工业场景对实时检测的需求。
To address the issues of insufficient detection accuracy and high model complexity in existing defect detec-tion algorithms for printed circuit board assembly components,this paper proposes a two-stage high-precision optical printed circuit board assembly(PCBA)component surface defect detection method based on target recognition and feature reconstruction.This method comprises two stages:component identification and defect detection.First,the component identification stage employs object detection algorithms to locate component regions on the PCBA.Through preprocessing and spatial overlap clipping,it reduces global background noise interference,ensuring semantic consistency of subsequent input features.This enhances defect detection accuracy and inference efficiency.In the second stage,an anomaly detection framework based on feature reconstruction is adopted,in which the pre-trained fea-ture extractor is enhanced to better handle random shapes and multi-scale defects while capturing richer channel and spatial information.To satisfy the real-time requirements of industrial inspection,a lightweight encoder with depth-wise separable convolutions is introduced to significantly reduce model complexity and improve generalization.Exper-imental results on both self-built PCBA datasets and the MvTec public dataset demonstrate that compared to existing state-of-the-art unsupervised defect detection algorithms,the proposed method achieves optimal performance in both image-level and pixel-level anomaly detection.Simultaneously,it significantly reduces model complexity,meeting the real-time detection requirements of industrial scenarios.
作者
何静飞
杨竹清
米成虎
李正龙
贺柳铭
HE Jingfei;YANG Zhuqing;MI Chenghu;LI Zhenglong;HE Liuming(School of Electronics and Information,Hebei University of Technology,Tianjin 300401,China;Hebei University of Tech-nology Innovation Research Institute(Shijiazhuang),Shijiazhuang,Hebei 050299,China)
出处
《河北工业大学学报》
2025年第6期79-88,共10页
Journal of Hebei University of Technology
基金
国家自然科学基金资助项目(62471174)
天津市自然科学基金资助项目(24JCYBJC00270)
石家庄市科技合作专项项目(SJZZXB23008)。
关键词
PCBA
缺陷检测
目标识别
特征重建
轻量级编码器
PCBA
defect detection
object detection
feature reconstruction
lightweight encoder
