计算机集成制造系统 ›› 2021, Vol. 27 ›› Issue (3): 716-727.DOI: 10.13196/j.cims.2021.03.006

• 当期目次 • 上一篇    下一篇

混合现实装配检测中深度学习的数据增强方法

王帅1,2,郭锐锋2+,董志勇3,王鸿亮2,张晓星1,2   

  1. 1.中国科学院大学计算机与控制学院
    2.中国科学院沈阳计算技术研究所
    3.陆军炮兵防空兵学院士官学校
  • 出版日期:2021-03-31 发布日期:2021-03-31
  • 基金资助:
    国家科技重大专项资助项目(2019ZX04014001-004);辽宁省重点研发计划资助项目(2018225096)。

Data enhancement method for deep learning in mixed reality assembly inspection

  • Online:2021-03-31 Published:2021-03-31
  • Supported by:
    Project supported by the National Science and Technology Major Project,China(No.2019ZX04014001-004),and the Liaoning Provincial Key Research and Development Program,China (No.2018225096).

摘要: 针对混合现实装配检测中,装配者的视觉检测位姿具有不确定性极易发生误检漏检的问题,提出一种混合现实装配检测中深度学习的数据增强方法。采用人为最佳数据增强策略的数据预处理方法,通过图像增强、几何变换、少量噪声干扰和随机遮挡的方式生成增强数据集,并改善图像增强过程中的特征失真,不仅能有效解决深度学习中人工标注样本任务量大的问题,还有助于提升检测模型的泛化能力。实验结果表明,该方法训练得到的新模型对汽车装配生产线零件的检测精度提升了11.38%。

关键词: 混合现实, 装配件检测, 深度学习, 数据增强

Abstract: Aiming at the problem that the assemblers visual inspection pose was uncertain and prone to error inspection and missed detection in the mixed reality assembly inspection,a data enhancement method for deep learning in mixed reality assembly inspection was proposed.Using an human-best data enhancement strategy,the data preprocessing method generated an enhanced data set by image enhancement,geometric transformation,small noise interference and random occlusion,and the feature distortion in the image enhancement process was weakened,which could not only effectively solve the problem of the large task of manual labeling sample in deep learning.It also helped to improve the generalization of the inspection model.The experimental results showed that the new model trained by the proposed method improved the inspection accuracy of parts of the automotive assembly line by 11.38%.

Key words: mixed reality, assembly inspection, deep learning, data enhancement

中图分类号: