摘要:水稻冠层信息在自动化管理上具有重要指导意义，但田间多变光照强度环境显著降低了水稻冠层图像分割和信息提取的精度。为降低光照强度的干扰，本文基于RGB、CIEL*a*b*、HSV色彩空间和多色彩空间（包括RGB、CIEL*a*b*和HSV色彩空间）构建水稻冠层图像的色彩特征组合，然后通过支持向量机（SVM）的线性核函数对水稻冠层图像进行分类识别，其分割方法分别定义为rgb-SVM、lab-SVM、hsv-SVM和Multi-SVM。同时，利用此方法对不同光照强度下的水稻冠层图像进行分割，并与常用的ExG&Otsu分割方法进行对比，比较不同方法的分割效果和光强稳健性。结果表明，rgb-SVM的分割效果优于ExG&Otsu方法，但对晴天条件下获取的水稻冠层图像的分割误差仍较大，光强稳健性低；lab-SVM和hsv-SVM分割方法的分割精确度较低，存在一定的欠分割现象；基于多色彩空间和支持向量机的Multi-SVM分割方法的分割效果最佳，该方法对不同光强下获取的水稻冠层图像的分割误差均控制在4.00%以内，具有较好的光强稳健性。因此，基于多色彩空间和支持向量机的Multi-SVM分割方法能够相对准确地将水稻像元从水稻冠层图像中分割出来，且对田间多变光强条件具有一定的稳健性，可为田间水稻生长发育监测和自动化管理提供一定的技术支持。
关键词:水稻/
冠层图像/
光照强度/
图像分割/
色彩空间/
支持向量机
Abstract:Digital image analysis of rice canopy has widely been used for monitoring rice growth, diagnosing rice nitrogen (N) content, controlling pests and predicting rice yield. But the accuracy, stability and reliability of digital image analysis of rice canopy has greatly relied on assumed segmentation precision of rice pixels. There is current a significant progress in auto-segmentation methods for plant images captured indoor or under controlled light conditions. However, it is still hard to segment images of rice canopy taken in outdoor environments with complex and changing illumination conditions. In this paper, we proposed a segmentation method for rice canopy images taken in outdoor environment that improves the accuracy and robustness of illumination of segmentation based on multi-color spaces and support vector machine (SVM) algorithm. The rice canopy images were taken using a digital camera (NikonD90, Nikon Inc., Tokyo, Japan) in August 11st to September 25th 2016 at the largest double-season rice production area in Pearl River Delta. The camera was mounted on a tripod at 1.5 m above rice canopy with straight downward looking posture. Three typical samples taken under different illumination conditions (which changed from sunny days to cloudy days and to overcast days) were treated as test images. The training data (including rice pixels and background pixels) for modeling the support vector machine classifier was randomly picked from the test images. The color features (r, g, b, L*, a*, b*, H, S, V) defined in 3 ordinarily used color spaces (RGB, CIEL*a*b* and HSV) of each pixel were calculated as training data. The SVM classifiers learned from the training data with the color features from RGB, CIEL*a*b*, HSV and multi-color spaces (including RGB, CIEL*a*b*, HSV) were defined as rgb-SVM, lab-SVM, hsv-SVM and Multi-SVM accordingly. The accuracy and robustness of the proposed methods were examined using the test images, which were next compared with ExG&Otsu (excess green index) performance. With the help of Photoshop image editing software, the ground-truth of the rice canopy images was labeled manually and treated as the reference for segmented error calculation, including false positive rate (the rate where segmentation algorithm falsely classed background pixels as rice pixels) and false negative rate (the rate that the segmented algorithm falsely classed the rice pixels as background pixels). The results showed that rgb-SVM algorithm performed better than ExG&Otsu algorithm. While segmentation errors of rgb-SVM algorithm for the images taken on overcast days and cloudy days were respectively 5.76% and 7.74%, that of rgb-SVM algorithm for the images taken on sunny days reached 16.99%. The accuracies of lab-SVM and hsv-SVM algorithms were unstable and high under-segmentation occurred under lab-SVM and hsv-SVM algorithms for images taken on cloudy days and sunny days. Multi-SVM algorithm had the best segmentation results, which were very close to ground-truth images. Specially, segmentation error of Multi-SVM algorithm for images taken on overcast days, cloudy days and sunny days were as low as 3.11%, 3.28% and 3.95%, respectively, which were lower than that for ExG&Otsu algorithm, especially for images taken on sunny days. The results showed that the accuracy of rice canopy extraction using Multi-SVM algorithm was significantly better than that using the other methods, particularly for images taken under high illumination conditions. The Multi-SVM algorithm based on multi-color spaces and support vector machine proposed in this paper accurately segmented and extracted rice pixels in rice canopy images. It was well-suited to the changing illumination in outdoor environment, thus providing valid data support for monitoring field rice growth under natural field conditions and automated rice farming.
Key words:Rice/
Canopy image/
Illumination condition/
Image segmentation/
Color space/
Support vector machine
PDF全文下载地址:
http://www.ecoagri.ac.cn/article/exportPdf?id=6491f3cc-b8c2-45cb-86a9-9bb4076ea3cb
