摘要
以云杉为研究对象提出了应用全卷积神经网络(Fully convolutional networks,FCN)分割图像的算法。利用无人机采集图像,标注470幅云杉图像,其中300幅组成训练集,170幅组成测试集,标注90幅樟子松图像作为附加测试集。以VGG16为基础建立云杉分割FCN模型,利用Tensorflow框架实现和训练网络,通过共享权值和逐渐降低的学习速率,提高FCN模型的训练性能。选择像素精度(PA)、均像素精度(MPA)、均交并比(MIoU)和频权交并比(FWIo U) 4个语义分割评价指标评价测试结果。FCN模型分割云杉图像,PA和MPA达到0. 86,MIoU达到0. 75,FWIo U达到0. 76,处理速率达到0. 085 s/幅,有效地解决了光照变化、云杉个体差异、地面杂草干扰和植株之间粘连的影响。与HSV颜色空间阈值分割以及K均值聚类分割算法比较,FCN模型的MIoU分别提高0. 10和0. 38。
Existing nursery inventory methods require people hand-counting, which is very labor consuming and not efficient. Using unmanned aerial vehicle (UAV) to facilitate counting the number of nursery-grown plants automatically with high accuracy provides an alternative to inventory management. The segmentation of individual plants in UAV images is the crucial step to achieve the plants counting task, which is challenging because of variations in illumination changes under natural conditions, the size difference between individual plants, the complicated background of the ground weeds and overlapping of adjacent plants. A spruce image segmentation algorithm based on fully convolutional networks (FCN) was proposed. Images were collected by using DIJ PHANTOM 4 in Inner Mongolia, in which 470 labeled spruce images with 300 images as training set, 170 images as test set, and 90 Pinus sylvestris images labeled as additional test set for comparing test results. To design FCN for accurate spruces segmentation, VGG16 was chosen as a basic network with the shared weights and the decreasing learning rate to improve the accuracy under Tensorflow framework. The results on the test set showed that FCN algorithm achieved effective spruces segmentation in spite of illumination changes, the size difference between individuals, the complicated background and the overlapping problem, with pixel accuracy (PA) of 0.86, mean pixel accuracy (MPA) of 0.86, mean intersection over union (MIoU) of 0.75 and frequency weighted intersection over union (FWIoU) of 0.76 at an average speed of 85 millisecond per image. Compared with K-means clustering segmentation algorithm and HSV threshold segmentation algorithm, the MIoU value of FCN algorithm was 0.10 and 0.38 higher, respectively. All of the test results showed that the proposed FCN algorithm provided an effective pipeline for plants segmentation.
作者
陈锋军
王成翰
顾梦梦
赵燕东
CHEN Fengjun;WANG Chenghan;GU Mengmeng;ZHAO Yandong(School of Technology, Beijing Forestry University, Beijing 100083, China;Beijing Laboratory of Urban and Rural Ecological Environment, Beijing Municipal Education Commission, Beijing 100083, China;Department of Horticultural Science, Texas A&M University, College Station TX 77843, USA;Key Laboratory of State Forestry Administration for Forestry Equipment and Automation, Beijing 100083, China)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2018年第12期188-194,210,共8页
Transactions of the Chinese Society for Agricultural Machinery
基金
中央高校基本科研业务费专项资金项目(2015ZCQ-GX-04)
国家重点研发计划项目(2017YFD0600901)
北京市科技计划项目(Z161100000916012)
北京市共建项目
关键词
云杉
图像分割
无人机
苗木库存统计
全卷积神经网络
spruce
image segmentation
unmanned aerial vehicle
nursery inventory statistics
fully convolutional networks
