IC器件表面缺陷多光谱图像特征融合检测方法  

Multispectral image feature fusion method for detecting surface defects in IC devices

作　　者：黄志海 邓耀华[1] 吴光栋 Huang Zhihai;Deng Yaohua;Wu Guangdong(School of Electromechanical Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学机电工程学院,广州510006

出　　处：《仪器仪表学报》2024年第9期24-33,共10页Chinese Journal of Scientific Instrument

基　　金：广东省基础与应用基础研究基金项目(2022B1515120053);广东省省级科技计划项目(2023A0505050151);东莞市重点领域研发项目(20221200300042)资助。

摘　　要：针对IC器件表面轻微缺陷在传统的像素级融合检测中容易被产生的冗余噪声淹没,干扰缺陷特征的提取,且在光照不稳定的复杂检测场景中不能够自适应调节可见光图像与红外图像对缺陷检测任务贡献度的问题,本文提出基于多光谱图像特征融合的IC器件表面缺陷检测方法,采用中期融合策略,设计了多光谱图像特征融合模块(MIFF),在YOLO框架下建立双路特征提取通道,构建多光谱图像特征融合端对端的YOLO-MIFF缺陷检测模型。实验表明,YOLO-MIFF融合检测比单可见光和单红外图像检测的mAP分别提高了24.69%、35.65%,相比于YOLO-Multiply、YOLO-Concat、YOLO-Add模型的检测精度分别提高了9.85%、6.67%、3.44%。To address the issue where minor surface defects of IC devices are often obscured by redundant noise in traditional pixel-level fusion detection—hindering defect feature extraction—and the challenge of adaptively adjusting the contribution of visible and infrared images in complex detection scenarios with unstable lighting,this paper proposes a surface defect detection method for IC devices based on multispectral image feature fusion.The method employs a mid-fusion strategy to design a Multispectral Image Feature Fusion(MIFF)module and establishes a dual-path feature extraction channel within the YOLO framework.This leads to the development of an end-to-end YOLO-MIFF defect detection model specifically for multispectral image feature fusion.Experimental results demonstrate that the YOLO-MIFF fusion detection model achieves a mean Average Precision(mAP)that is 24.69%and 35.65%higher than that of single visible and single infrared image detection,respectively.Additionally,compared to the YOLO-Multiply,YOLO-Concat,and YOLO-Add models,YOLO-MIFF improves detection accuracy by 9.85%,6.67%,and 3.44%,respectively.

关 键 词：IC器件 缺陷检测 多光谱图像 深度学习 

分 类 号：TH391.4[机械工程—机械制造及自动化] TH89