当前位置: 首页 > 文章 > 基于改进Mask RCNN的复杂环境下苹果检测研究 中国农机化学报 2019 (10) 128-134
Position: Home > Articles > Research on apple detection in complex environment based on improved Mask RCNN Journal of Chinese Agricultural Mechanization 2019 (10) 128-134

基于改进Mask RCNN的复杂环境下苹果检测研究

作  者:
岳有军;田博凯;王红君;赵辉
单  位:
天津理工大学天津市复杂系统控制理论与应用重点实验室;天津农学院
关键词:
苹果检测;深度学习;采摘机器人;机器视觉
摘  要:
苹果检测是苹果采摘系统中的关键环节,为实现复杂环境下苹果采摘机器人视觉系统对苹果的识别和定位,提出一种基于深度学习的方法,通过改进的Mask RCNN网络对苹果进行检测研究。该方法在原始Mask RCNN网络的基础上,增加边界加权损失函数,能够使边界检测结果更为精确。训练后的模型在验证集下的AP值为92.62%。通过比较Mask RCNN与Faster RCNN、YOLO v3和传统分割算法K-means算法在不同数目,不同光照和绿色苹果情况下的检测效果,试验结果表明:Mask-RCNN的F1值和分割效果均高于其他算法,证明本文方法对复杂环境下的苹果有很好的检测效果,可为苹果产业中采摘机器人的视觉系统提供技术支持。
译  名:
Research on apple detection in complex environment based on improved Mask RCNN
作  者:
Yue Youjun;Tian Bokai;Wang Hongjun;Zhao Hui;Tianjin Key Laboratory of Complex System Control Theory and Applications, Tianjin University of Technology;Tianjin Agricultural College;
关键词:
apple detection;;deep learning;;picking robot;;machine vision
摘  要:
Apple detection is a crucial link in the apple picking system. In order to realize the apple picking robot vision system's identification and location of apples in a complex environment, this paper proposes a deep learning-based method to detect apples through improved Mask RCNN network. Based on the original Mask RCNN network, this method adds a boundary weight loss function, which can make the boundary detection result more accurate. The trained model has an AP value of 92.62% under the verification set. By comparing the detection effects of Mask RCNN, Faster RCNN, YOLO v3 and traditional segmentation algorithm K-means in different numbers, different illumination and green apple, the experimental results show that the F1 value and segmentation effect of Mask-RCNN are higher than other algorithms. The result proves that this method has a good detection effect on apples in complex environments, and it can provide technical support for the vision system of picking robots in the apple industry.

相似文章

计量
文章访问数: 22
HTML全文浏览量: 0
PDF下载量: 0

所属期刊

推荐期刊