摘要
为了使采摘机器人在收获番茄时更加精准地识别目标果实,采用改进后的Cascade RCNN网络对温室内的番茄果实进行目标检测。将Cascade RCNN网络中的非极大值抑制算法替换为Soft-NMS(soft non-maximum suppression)算法,采用适合番茄形状的锚框,增强网络对重叠果实的识别能力,与原Cascade RCNN网络相比,目标识别的准确率提高了近2%,在识别番茄果实的同时,该网络对番茄的成熟度进行了简单分类。为进一步验证网络性能,将改进网络与经典的Faster RCNN网络和YOLOv3网络进行对比。实验结果表明,改进网络能够准确地识别出番茄果实,并对成熟番茄与未成熟番茄做出区分。该方法可为温室内番茄果实的采摘提供技术支持。
To recognize the target fruit in harvest more accurately for robot picking,,an improved Cascade RCNN network was used to detect tomato fruit maturity in greenhouse.In this method,the non-maximum suppression algorithm in Cascade RCNN network was replaced by the Soft-NMS algorithm,and an anchor frame suitable for tomato shape was adopted,by which the network's ability to recognize overlapped fruits was enhanced.Compared with the original Cascade RCNN network,the accuracy of target recognition was improved by approximately 2%.While identifying tomato fruits,the network performed a simple classification of tomato maturity.To further verify the network performance,the improved network was compared with the classical Faster RCNN network and the YOLOv3 network.Results show that the improved network could accurately identify tomato fruit maturity for harvest and distinguish between mature tomato and immature tomato.This method provides a technical support for robot tomato harvest in greenhouse.
作者
岳有军
孙碧玉
王红君
赵辉
YUE You-jun;SUN Bi-yu;WANG Hong-jun;ZHAO Hui(Tianjin Key Laboratory of Complex System Control Theory and Applications,School of Electrical and Electronic Engineering,Tianjin University of Technology,Tianjin 300384,China;College of Engineering and Technology,Tianjin Agricultural College,Tianjin 300392,China)
出处
《科学技术与工程》
北大核心
2021年第6期2387-2391,共5页
Science Technology and Engineering
基金
天津市科技支撑计划(17ZXYENC00080,18YFZCNC01120,15ZXZNGX00290)。
作者简介
第一作者:岳有军(1970—),男,汉族,天津人,博士,教授。研究方向:复杂系统建模及智能控制、机器人导航与控制技术、电力电子技术及应用。E-mail:1547986714@qq.com。
