基于YOLOv5的芯片表面缺陷检测算法优化
Optimization of Chip Surface Defect Detection Algorithm Based on YOLOv5
王洋 1梁礼明1
作者信息
- 1. 江西理工大学电气工程与自动化学院,江西赣州 341000
- 折叠
摘要
芯片工艺中存在许多不确定因素,导致其缺陷类型众多,且缺陷特征难以定义.传统的缺陷检测方法容易出现漏检、误检及准确性欠佳等弊端,为此,提出一种以YOLOv5 为基础的新型集成电路表面缺陷检测算法.首先,在YOLOv5 网络结构中融入注意力机制,以更好地识别芯片表面缺陷;然后,引入较为复杂双向融合网络BiFPN与ASFF,得到 4 种改进模型;最后,将 4 种模型进行对比,以获得芯片表面缺陷检测的最优模型.与原始YOLOV5 模型相比,4 种新型网络模型均展示出显著提升.其中,使用CBAM注意力机制和BiFPN特征融合网络模型的mAP获得了0.723的最大值.
Abstract
There are many uncertain factors in chip manufacturing,resulting in numerous types of defects and difficulty in defining defect characteristics.Traditional defect detection methods are prone to defects such as missed detections,false detections,and poor accuracy.Therefore,a new integrated circuit surface defect detection algorithm based on YOLOv5 is proposed.Firstly,incorporating attention mechanism into the YOLOv5 network structure to better identify chip surface defects.Then,four improved models were obtained by introducing complex bidirectional fusion networks BiFPN and ASFF.Finally,compare the models to obtain the optimal model for chip surface defect detection.Compared with the original YOLOv5 model,all four new network models show significant improvements.Among them,the mAP obtained a maximum value of 0.723 using CBAM attention mechanism and BiFPN feature fusion network model.
关键词
YOLOv5/非极大值抑制/BiFPN/ASFFKey words
YOLOv5/non-maximum suppression/BiFPN/ASFF引用本文复制引用
出版年
2024
NETL
NSTL
万方数据