第16卷第3期 智能系统学报 Vol.16 No.3 2021年5月 CAAI Transactions on Intelligent Systems May 2021 D0L:10.11992tis.202005016 改进Center-Net网络的自主喷涂机器人室内窗户检测 洪恺临'，曹江涛'，姬晓飞2 (1.过宁石油化工大学信息与控制工程学院，辽宁抚顺113001；2.沈阳航空航天大学自动化学院，辽宁沈阳 110136) 摘要：室内自主喷涂机器人可以实现室内墙面喷涂的自动化以此提升喷涂的效率，诚少人力物力的投入。而 基于计算机视觉的室内窗户检测算法则是该机器人的关键技术。对于室内窗户检测，由于环境光照、窗户形状 和窗户透光属性的存在，传统方法无法得到较好的效果。针对此问题，设计一种基于深度学习的室内窗户检测 算法。该算法主要对中心点网络(CenterNet))的特征提取网络进行修改，减少部分卷积操作，使用Ghost模块替 换原始的卷积模块，降低特征冗余，并引入注意力机制，让网络尽可能表达重要信息。实验结果表明，改进的 CenterNet在不损失网络精度的前提下，大幅度提高了网络的运算速度，使得该检测算法即使在机器人端的嵌 入式系统上也可以达到实时检测的效果。 关键词：喷涂机器人；深度学习；目标检测；室内窗户检测；中心点网络；Gost模块；注意力机制：嵌入式设备 中图分类号：TP391.1文献标志码：A文章编号：1673-4785(2021)03-0425-08 中文引用格式：洪恺临，曹江涛，姬晓飞.改进Center-Net网络的自主喷涂机器人室内窗户检测IJ.智能系统学报，2021， 16(3):425-432. 英文引用格式：HONG Kailin,CAO Jiangtao,JI Xiaofei..Indoor window detection of autonomous spraying robot based on im- proved CenterNet networkJ).CAAI transactions on intelligent systems,2021,16(3):425-432. Indoor window detection of autonomous spraying robot based on improved CenterNet network HONG Kailin',CAO Jiangtao',JI Xiaofei? (1.School of Information and Control Engineering,Liaoning Shihua University,Fushun 113001,China;2.School of Automation, Shenyang Aerospace University,Shenyang 110136,China) Abstract:An indoor autonomous spraying robot can realize the automation of indoor wall spraying to improve the effi- ciency of spraying and reduce the investment of manpower and material resources.The indoor window detection al- gorithm based on computer vision is the key technology of the robot.For indoor window detection,traditional methods cannot obtain good results owing to the actual scene's requirements for recognition speed and accuracy as well as the presence of lighting in the environment,shape of the window,and light transmission properties of the window.To solve this problem,an indoor window detection algorithm based on deep learning is designed.This algorithm mainly modifies the backbone feature extraction of the CenterNet network,reduces part of the convolution operation,replaces the origin- al convolution module with ghost block,reduces the redundancy feature,and introduces an attention mechanism to keep the network under a limited number of parameters that express important information as much as possible.The experi- mental results show that the improved CenterNet algorithm greatly improves the operation speed of the network without losing the accuracy of the network so that the network can achieve a real-time detection effect even on the embedded system of the robot. Keywords:spraying robot;deep learning;target detection;indoor window detection;Center-Net;Ghost block;atten- tion mechanism;embedded device 随着城镇化水平的不断提高，室内装修领域大部分的室内喷涂工作仍然是以人工喷涂为主， 的自动化发展水平受到越来越多人的关注。目前 喷涂的效率低且质量难以保证。现有的室内喷涂 收稿日期：2020-05-12. 机器人虽然可以进行简单的墙面喷涂，但是都缺 基金项目：国家自然科学基金项目(61673199)：辽宁省科技公 少环境中不可喷涂区域的识别，如果希望喷涂机 益研究基金项目(2016002006). 通信作者：姬晓飞.E-mail:jixiaofei7804@126.com 器人真正地做到自主喷涂，那么对于窗户的检测DOI: 10.11992/tis.202005016 改进 Center-Net 网络的自主喷涂机器人室内窗户检测 洪恺临1 ，曹江涛1 ，姬晓飞2 （1. 辽宁石油化工大学 信息与控制工程学院，辽宁 抚顺 113001; 2. 沈阳航空航天大学 自动化学院，辽宁 沈阳 110136） 摘 要：室内自主喷涂机器人可以实现室内墙面喷涂的自动化以此提升喷涂的效率，减少人力物力的投入。而 基于计算机视觉的室内窗户检测算法则是该机器人的关键技术。对于室内窗户检测，由于环境光照、窗户形状 和窗户透光属性的存在，传统方法无法得到较好的效果。针对此问题，设计一种基于深度学习的室内窗户检测 算法。该算法主要对中心点网络 (CenterNet) 的特征提取网络进行修改，减少部分卷积操作，使用 Ghost 模块替 换原始的卷积模块，降低特征冗余，并引入注意力机制，让网络尽可能表达重要信息。实验结果表明，改进的 CenterNet 在不损失网络精度的前提下，大幅度提高了网络的运算速度，使得该检测算法即使在机器人端的嵌 入式系统上也可以达到实时检测的效果。 关键词：喷涂机器人；深度学习；目标检测；室内窗户检测；中心点网络；Ghost 模块；注意力机制；嵌入式设备 中图分类号：TP391.1 文献标志码：A 文章编号：1673−4785(2021)03−0425−08 中文引用格式：洪恺临, 曹江涛, 姬晓飞. 改进 Center-Net 网络的自主喷涂机器人室内窗户检测 [J]. 智能系统学报, 2021, 16(3): 425–432. 英文引用格式：HONG Kailin, CAO Jiangtao, JI Xiaofei. Indoor window detection of autonomous spraying robot based on im￾proved CenterNet network[J]. CAAI transactions on intelligent systems, 2021, 16(3): 425–432. Indoor window detection of autonomous spraying robot based on improved CenterNet network HONG Kailin1 ，CAO Jiangtao1 ，JI Xiaofei2 (1. School of Information and Control Engineering, Liaoning Shihua University, Fushun 113001, China; 2. School of Automation, Shenyang Aerospace University, Shenyang 110136, China) Abstract: An indoor autonomous spraying robot can realize the automation of indoor wall spraying to improve the effi￾ciency of spraying and reduce the investment of manpower and material resources. The indoor window detection al￾gorithm based on computer vision is the key technology of the robot. For indoor window detection, traditional methods cannot obtain good results owing to the actual scene’s requirements for recognition speed and accuracy as well as the presence of lighting in the environment, shape of the window, and light transmission properties of the window. To solve this problem, an indoor window detection algorithm based on deep learning is designed. This algorithm mainly modifies the backbone feature extraction of the CenterNet network, reduces part of the convolution operation, replaces the origin￾al convolution module with ghost block, reduces the redundancy feature, and introduces an attention mechanism to keep the network under a limited number of parameters that express important information as much as possible. The experi￾mental results show that the improved CenterNet algorithm greatly improves the operation speed of the network without losing the accuracy of the network so that the network can achieve a real-time detection effect even on the embedded system of the robot. Keywords: spraying robot; deep learning; target detection; indoor window detection; Center-Net; Ghost block; atten￾tion mechanism; embedded device 随着城镇化水平的不断提高，室内装修领域 的自动化发展水平受到越来越多人的关注。目前 大部分的室内喷涂工作仍然是以人工喷涂为主， 喷涂的效率低且质量难以保证。现有的室内喷涂 机器人虽然可以进行简单的墙面喷涂，但是都缺 少环境中不可喷涂区域的识别，如果希望喷涂机 器人真正地做到自主喷涂，那么对于窗户的检测 收稿日期：2020−05−12. 基金项目：国家自然科学基金项目 (61673199)；辽宁省科技公 益研究基金项目 (2016002006). 通信作者：姬晓飞. E-mail：jixiaofei7804@126.com. 第 16 卷第 3 期 智 能 系 统 学 报 Vol.16 No.3 2021 年 5 月 CAAI Transactions on Intelligent Systems May 2021