基于M-LP特征加权聚类的果树冠层图像分割方法  

Fruit Tree Canopy Image Segmentation Method Based on M-LP Features Weighted Clustering

作　　者：程浈浈 祁力钧[1] 程一帆 吴亚垒 张豪 肖雨 CHENG Zhenzhen;QI Lijun;CHENG Yifan;WU Yalei;ZHANG Hao;XIAO Yu(College of Engineering,China Agricultural University,Beijing 100083,China;Beijing Siweitu New Technology Co.,Ltd.,Beijing 100094,China)

机构地区：[1]中国农业大学工学院,北京100083 [2]北京四维图新科技股份有限公司,北京100094

出　　处：《农业机械学报》2020年第4期191-198,260,共9页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家重点研发计划项目(2017YFD0701400、2016YFD0200700)。

摘　　要：针对背景和杂草干扰下的果树图像冠层提取问题,提出了一种基于M-SP特征加权聚类的冠层分割算法。首先,将采集的原始图像由RGB颜色空间转换到HSI颜色空间,计算果树与背景区域在H、S分量上的马氏距离,构造马氏距离相似度矩阵M;其次,提取图像像素的垂直位置作为空间特征P,在HSI空间内的I分量上,利用最大熵算法提取图像的阴影区域,并进行掩膜处理,将获取的阴影区域作为空间特征的加权区域L,从而构造阴影位置加权的空间特征LP;最后,对获取的M-LP特征矩阵进行归一化处理,分别进行上背景、下背景、果树冠层、杂草4个类别的K-means聚类,最终完成图像分割。为验证算法的有效性,在采集的果树图像上进行了分割试验,结果表明,基于M-LP特征的聚类方法能有效解决重度杂草干扰条件下果树冠层被漏分的问题。采用精确率、召回率和F1值3个评价指标对分割结果进行定量评价,选取不同杂草干扰程度(轻微、中等、较强)和时间段(早晨、中午、傍晚)的果树图像,分别以传统K-means和GMM聚类算法作为对比进行试验,结果表明,相对于未经过特征提取的普通聚类分割方法,本文算法对于不同杂草干扰程度和不同拍摄时间段下的果树冠层分割表现出一定的鲁棒性,平均精确率为87.1%,平均召回率为87.7%,平均F1值为87.1%。分割和验证结果表明,在进行有效图像特征提取的基础上,结合少量标注作为先验知识的无监督分割方法可以准确分割出果树冠层区域。Fruit canopy information collection plays an important role in the orchard variable spray.Aiming at the problem of canopy extraction of fruit trees under background( non-green plants) and weed disturbance,a canopy segmentation algorithm was proposed based on M-SP feature weighted clustering.The segmentation process can be described as: the original image was converted from RGB color space to HSI color space. The Mahalanobis distance similarity matrix( M) was constructed by calculating the hue( H) and saturation( S) components between fruit tree and the background;moreover,luminance feature( L) was extracted: the vertical position of the pixel was used as the position feature( P). The maximum entropy algorithm was used to extract the shadow region of the image and perform mask processing on the intensity( I) component in the HSI. The obtained shadow region was used as weighted region S of spatial feature,thereby constructing the shadow position weighting. Finally,the acquired M-SP feature matrix was normalized,and the K-means clustering of the upper background,the lower background,the fruit canopy and the weeds were respectively performed,and the image segmentation was finally completed. In order to verify the accuracy of the quantitative verification algorithm,precision,recall and F1-scores were used to evaluate the image segmentation results under different weed disturbance levels( slight,medium and strong) and time segments( morning,noon and evening). The K-means and Gaussian mixture model( GMM) without feature extraction were used respectively as comparative experiments. The results showed that the proposed method was robust to the canopy segmentation of fruit trees under different weed interference levels and different shooting time periods. The average precision was 87. 1%,the average recall was 87. 7%,and the average F1-scores was 87. 1%. The segmentation and verification results showed that the algorithm can accurately segment the canopy area of fruit trees,which provided a reference for collecting the can

关 键 词：果树冠层 图像分割 杂草干扰 马氏距离 特征提取 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]