摘要
现有汽车零部件表面缺陷检测方法大多数都是依靠人工目检或传统的图像处理方法,其检测精度和速度都不能满足零部件工厂需求。由于汽车零部件的残次率低,导致可用的数据量少,一般的深度学习模型不能很好地应用于汽车零部件表面缺陷检测。针对上述问题,提出一种基于机器视觉的小样本汽车零部件表面缺陷检测方法。上述方法在Faster RCNN检测网络基础上,采用指导框区域候选网络改进原有的区域候选网络,并且利用聚焦式损失函数来进一步改善正负样本不均衡的问题,同时加入循环特征金字塔结构以及组合特征关系检测器。在汽车零部件表面缺陷数据集和小样本FSOD数据集上的实验结果表明,小样本汽车零部件表面缺陷检测模型较好地实现了在小样本零部件数据条件下对零部件表面缺陷的检测。
Most of the existing surface defect detection methods of automobile parts rely on manual visual inspection or traditional image processing methods,and their detection accuracy and speed can not meet the needs of parts factories.Due to the low defect rate of auto parts,the amount of available data is small,and the general deep learning model can not be well applied to the surface defect detection of auto parts.To solve the above problems,a few-shot surface defect detection method of automobile parts based on machine vision is proposed.Based on the Faster RCNN detection network,this method uses the guided anchoring region proposal network to improve the original region proposal network,and uses the focal loss function to further improve the imbalance between positive and negative samples.At the same time,it adds the recursive feature pyramid structure and the combined feature relation detector.The experimental results on automobile parts surface defect data set and few-shot FSOD data set show that the few-shot automobile parts surface defect detection model can better detect parts surface defects under the condition of few-shot parts data.
作者
佟鑫
郑彤
于重重
叶洋
TONG Xin;ZHENG Tong;YU Chong-chong;YE Yang(College of Artificial Intelligence,Beijing Technology and Business University,Beijing100048,China)
出处
《计算机仿真》
北大核心
2023年第4期160-164,212,共6页
Computer Simulation
基金
北京市自然科学基金资助项目(4202015)。
