摘要
芯片的表面缺陷检测在半导体制造中具有重要意义,针对目前芯片表面缺陷面积小,缺陷外形多变,缺陷尺寸跨度大的情况,提出一种基于YOLOv5改进的芯片表面缺陷检测算法,首先基于ConvNext网络改进特征提取模块,提升网络稳定性和特征表达能力,同时提出增强卷积注意力模块(ehanced convolutional block attention module,E_CBAM),将更详细的位置信息嵌入到卷积注意力(convolutional block attention module,CBAM)之中,提升整个网络对于小面积及边缘缺陷的检测能力,而针对芯片缺陷多变尺寸跨度大的问题,研究引入了可变形卷积和双向特征金字塔网络(bi-directional feature pyramid network,BiFPN),一方面可变形卷积对于外形不规则的卷积有更好的提取能力,另一方面Neck部分的BiFPN在简化结构的同时保证了多尺度融合的准确性。经过实验表明,改进后的网络在芯片表面缺陷数据集(chip defect dataset,CDD)上,平均精度均值(mAP)mAP@0.5指标达到95.3%,相较于原始的YOLOv5s网络提升了3.1%,在没有过多增加网络参数的情况下,对芯片表面缺陷的精度更高,鲁棒性更强。
Chip surface defect detection is of great significance in semiconductor manufacturing,for the current chip surface defect area is small,defect shape is variable,defect size spanning a large situation,put forward an improved chip surface defect detection algorithm based on YOLOv5,first of all,based on the CnovNext network to improve the feature extraction module,improve the stability of the network and the ability of feature expression,and at the same time put forward the ehanced convolutional block attention module(E_CBAM)module is proposed to embed more detailed location information into convolutional block attention module(CBAM)to improve the detection capability of the whole network for small area and edge defects.For the problem of large size span of chip defects,the study introduces deformable convolution and BiFPN module,on the one hand,the deformable convolution has better extraction ability for irregular shape convolution,on the other hand,the BiFPN in the Neck part simplifies the structure and ensures the accuracy of multi-scale fusion.After the experiments,it is shown that the improved network achieves a mAP@0.5 index of 95.3%on chip surface defect dataset(CDD),which is 3.1%higher compared to the original YOLOv5s network,which is more accurate and robust to the chip surface defects without too much increase in the network parameters.
作者
韩明桥
蒋三新
Han Mingqiao;Jiang Sanxin(College of Electronics and Information Engineering,Shanghai University of Electric Power,Shanghai 201306,China)
出处
《国外电子测量技术》
2024年第1期45-51,共7页
Foreign Electronic Measurement Technology
作者简介
韩明桥,硕士,主要研究方向为图像处理与识别、缺陷检测技术等。E-mail:1136770381@qq.com;通信作者:蒋三新,博士,讲师,主要研究方向为机器视觉、深度增强学习等。E-mail:samjoe_2018@shiep.edu.cn。
