第16卷第6期 智能系统学报 Vol.16 No.6 2021年11月 CAAI Transactions on Intelligent Systems Now.2021 D0:10.11992/tis.202011003 网络出版地址：https:/kns.cnki.net/kcms/detail/23.1538.TP.20210330.1755.006.html 改进Faster R-CNN的汽车仪表指针实时检测 伍锡如，邱涛涛 (桂林电子科技大学电子工程与自动化学院，广西桂林541004) 摘要：针对产业化的汽车仪表指针人工视觉检测效果差、检测速度慢和实时性低等问题，本文提出了一种改 进的Faster R-CNN汽车仪表指针实时检测算法。通过改进原始的RolI网络层结构，实现小目标高低层特征之 间的完整传递：采用双线性内插算法替代两次量化操作，使得特征聚集变成连续的过程，能够有效减少计算时 间：最后将工业机采集的视频数据，预处理成VOC格式数据集进行训练，调整超参数得到改进汽车仪表指针检 测模型。实验结果表明：所提出的方法能够快速、准确地实现汽车仪表指针检测，单张图片的平均检测时间为 0.197s,平均检测精度可达92.7%。在不同类别仪表指针的迁移实验中，展示了良好的泛化性能。 关键词：卷积神经网络；汽车仪表指针；实时检测：双线性内插；深度学习；模式识别：特征提取；特征聚集 中图分类号：TP183:TP391.41文献标志码：A文章编号：1673-47852021)06-1056-08 中文引用格式：伍锡如，邱涛涛.改进Faster R-CNN的汽车仪表指针实时检测JL.智能系统学报，2021,16(6)：1056-1063. 英文引用格式：WU Xiru,,QIU Taotao.Improved Faster R-CNN vehicle instrument pointer real--time detection algorithm.CAAl transactions on intelligent systems,2021,16(6):1056-1063. Improved Faster R-CNN vehicle instrument pointer real-time detection algorithm WU Xiru,QIU Taotao (College of Electronic Engineering and Automation,Guilin University of Electronic Technology,Guilin 541004,China) Abstract:This paper proposes an improved Faster R-CNN vehicle instrument pointer real-time detection algorithm to solve the problems of the poor artificial visual detection effect,slow detection speed,and low real-time performance of industrialized vehicle instrument pointers.First,complete transfer between the high-and low-layer features of a small target is realized by improving an original Rol network layer structure.Subsequently,continuous feature aggregation re- duces calculation time using a bilinear interpolation algorithm to replace two quantization operations.Finally,video data collected by an industrial machine are preprocessed into a VOC format data set for training,and hyperparameters are ad- justed to obtain an improved vehicle instrument pointer detection model.Experimental results show that the proposed method can quickly and accurately detect the vehicle instrument pointer.The average detection time of a single picture is 0.197s,and the average detection accuracy can reach 92.7%.The good generalization performance of this method is demonstrated in the migration experiment of different instrument pointer types. Keywords:convolutional neural network;vehicle instrument pointer;real-time detection;bilinear interpolation;deep learning;pattern recognition;feature extraction;feature aggregation 随着汽车仪表生产过程的自动化需求不断提动检测技术不断激增的需求4，大多数企业仍采 高，如何开发出一种高效、实时的指针检测方法 用人工视觉视检的方法来检测仪表指针，人工检 成为当前人工智能领域的热点课题。面对自 测方法受工作状态影响，导致检测标准难以统 一，无法满足流水线上仪表的产量需求，检测效 收稿日期：2020-11-03.网络出版日期：2021-03-31. 基金项目：国家自然科学基金项目(61863007)：广西自然科学 率和精度也随之受到影响。由于指针目标较小， 基金项目(2020 GXNSFDA238029):广西研究生教有 创新计划项目(YCSW2020159). 对应的像素中包含的特征很少，会导致存在漏检 通信作者：邱涛涛.E-mail:18339171275@163.com 的情况。DOI: 10.11992/tis.202011003 网络出版地址: https://kns.cnki.net/kcms/detail/23.1538.TP.20210330.1755.006.html 改进 Faster R-CNN 的汽车仪表指针实时检测 伍锡如，邱涛涛 （桂林电子科技大学 电子工程与自动化学院，广西 桂林 541004） 摘 要：针对产业化的汽车仪表指针人工视觉检测效果差、检测速度慢和实时性低等问题，本文提出了一种改 进的 Faster R-CNN 汽车仪表指针实时检测算法。通过改进原始的 RoI 网络层结构，实现小目标高低层特征之 间的完整传递；采用双线性内插算法替代两次量化操作，使得特征聚集变成连续的过程，能够有效减少计算时 间；最后将工业机采集的视频数据，预处理成 VOC 格式数据集进行训练，调整超参数得到改进汽车仪表指针检 测模型。实验结果表明：所提出的方法能够快速、准确地实现汽车仪表指针检测，单张图片的平均检测时间为 0.197 s，平均检测精度可达 92.7%。在不同类别仪表指针的迁移实验中，展示了良好的泛化性能。 关键词：卷积神经网络；汽车仪表指针；实时检测；双线性内插；深度学习；模式识别；特征提取；特征聚集 中图分类号：TP183; TP391.41 文献标志码：A 文章编号：1673−4785(2021)06−1056−08 中文引用格式：伍锡如, 邱涛涛. 改进 Faster R-CNN 的汽车仪表指针实时检测 [J]. 智能系统学报, 2021, 16(6): 1056–1063. 英文引用格式：WU Xiru, QIU Taotao. Improved Faster R-CNN vehicle instrument pointer real-time detection algorithm[J]. CAAI transactions on intelligent systems, 2021, 16(6): 1056–1063. Improved Faster R-CNN vehicle instrument pointer real-time detection algorithm WU Xiru，QIU Taotao (College of Electronic Engineering and Automation, Guilin University of Electronic Technology, Guilin 541004, China) Abstract: This paper proposes an improved Faster R-CNN vehicle instrument pointer real-time detection algorithm to solve the problems of the poor artificial visual detection effect, slow detection speed, and low real-time performance of industrialized vehicle instrument pointers. First, complete transfer between the high- and low-layer features of a small target is realized by improving an original RoI network layer structure. Subsequently, continuous feature aggregation re￾duces calculation time using a bilinear interpolation algorithm to replace two quantization operations. Finally, video data collected by an industrial machine are preprocessed into a VOC format data set for training, and hyperparameters are ad￾justed to obtain an improved vehicle instrument pointer detection model. Experimental results show that the proposed method can quickly and accurately detect the vehicle instrument pointer. The average detection time of a single picture is 0.197 s, and the average detection accuracy can reach 92.7%. The good generalization performance of this method is demonstrated in the migration experiment of different instrument pointer types. Keywords: convolutional neural network; vehicle instrument pointer; real-time detection; bilinear interpolation; deep learning; pattern recognition; feature extraction; feature aggregation 随着汽车仪表生产过程的自动化需求不断提 高，如何开发出一种高效、实时的指针检测方法 成为当前人工智能领域的热点课题[1-3]。面对自 动检测技术不断激增的需求[4-5] ，大多数企业仍采 用人工视觉视检的方法来检测仪表指针，人工检 测方法受工作状态影响，导致检测标准难以统 一，无法满足流水线上仪表的产量需求，检测效 率和精度也随之受到影响。由于指针目标较小， 对应的像素中包含的特征很少，会导致存在漏检 的情况。 收稿日期：2020−11−03. 网络出版日期：2021−03−31. 基金项目：国家自然科学基金项目 (61863007)；广西自然科学 基金项目 (2020GXNSFDA238029)；广西研究生教育 创新计划项目 (YCSW2020159). 通信作者：邱涛涛. E-mail：18339171275@163.com. 第 16 卷第 6 期 智 能 系 统 学 报 Vol.16 No.6 2021 年 11 月 CAAI Transactions on Intelligent Systems Nov. 2021