摘要
关键零件完整性是燃气表重要检定要求之一,经典图像特征匹配方法实现其完整性检测,存在通用性、泛化能力较低问题。本文提出一种改进Faster R-CNN多视角燃气表关键零件识别定位方法,该方法首先采用Vision Transformer(ViT)替代Faster R-CNN卷积神经网络,其自注意力机制促进学习图像块特征之间相关性,强化表征能力;其次研究ViT优化结构参数,在Transformer层数L=14、自注意头数m=12下,模型可达到相对较优准确率。实验表明,最优模型mAP达86.71%,较ResNet50提高2.48%,与ResNet101检测准确率相当,能有效降低模型复杂性,检测效率提高5.8%;燃气表关键零件单次检测耗时1.13 s,可满足燃气表外观关键零件检测的准确性、实时性要求。
The completeness of key parts is an important verification requirement for gas meters.Although the traditional image feature matching method is used to realize the automation of part detection,its universality is poor.This paper proposes an improved method for Faster R-CNN to identify and locate key parts of gas meters from multiple perspectives.First,Faster R-CNN utilizes Vision Transformer(ViT)to replace the convolutional neural networks,whose self-attention mechanism can help to learn the correlation between image block features and strengthen the representation ability.And then the ViT structure with 14 Transformer layers and 12 self-attention heads is optimized to achieve optimal accuracy.Experimental results show that the mAP of the optimal model is 86.71%,2.48%higher than that of ResNet50o.It is equivalent to the detection accuracy of ResNetlol,whose detection efficiency is increased by 5.8%,and effectively reduces the complexity of the model.It takes 1.13 s to accomplish the single detection of key parts of gas meter.The method balances the accuracy and real-time ability for key parts detection of gas meter.
作者
高泽铭
刘桂雄
陈国宇
黄坚
Gao Zeming;Liu Guixiong;Chen Guoyu;Huang Jian(School of Mechanical and Automotive Engineering,South China University of Technology,Guangzhou 510640,China;Guangzhou Institute of Energy Testing,Guangzhou 511447,China;Guangzhou Institute of Measurement and Testing Technology,Guangzhou 510663,China)
出处
《电子测量技术》
北大核心
2023年第11期7-12,共6页
Electronic Measurement Technology
基金
广东省市场监督管理局科技项目(2022CJ04)
广东省市场监督管理局科技项目(XMBH20220614019)资助。
