Abstract: In order to resolve the difficult problem of rice canopy image segmentation caused by variable light intensities in the fields, this paper discussed a rice canopy image segmentation algorithm based on support vector machine (SVM) and maximum variance between classes (OTSU). Firstly, 2 class training images (rice S1 and background S2) were captured from the rice canopy images under different illumination conditions, and the density of R, G, B, r, g, b, L*, a*, b*, H, S, and V color indices from RGB, rgb(standard RGB), CIE L*a*b* and HSV color spaces for rice (S1) and background (S2) training images were analyzed. The g, a*, b*, and S color indices, whose histogram distributions of S1 and S2 were significant bimodal, were selected as key color indices. Afterwards, an optimal hyperplane Z（Z=0.421g+0.753a*+0.152b*+0.051S+0.085） in multi\|dimensional color spaces including g, a*, b*, and S color indices were obtained by support vector machine. Finally, the Z(x, y) of any rice canopy image was calculated, and threshold Zt value was calculated by OTSU method. The accuracy and robustness of the proposed method was examined over 90 test images under different illumination conditions (overcast days, cloudy days and sunny days), and compared with the performances of ExG & OTSU. The accuracy of segmentation and light intensity robustness of this method were analyzed. The results showed that the accuracy of ExG & OTSU method was influenced significantly by light intensity, and decreased along with the light intensity increasing. The robustness of light intensity was poor. The segmented error of the proposed method for the images, taken on overcast days, cloudy days, and sunny days were 7.30%, 8.72%, and 8.98%, respectively. The accuracy of segmentation was higher and with better robustness of light intensity. Therefore, the proposed method had higher segmentation accuracy and robustness of light intensity, and could provide technical reference for accurate segmentation of rice canopy image in fields under changeable illumination.

Key words: support vector machine, OTSU, color space, rice, light intensity, image segmentation

摘要： 针对田间多变光照强度给水稻冠层图像分割带来的难题，探讨了一种基于支持向量机（SVM）和最大类间方差法（OTSU法）相结合的水稻冠层图像分割算法。首先，从不同光强条件下的水稻冠层图像中采集代表性水稻和背景像元构建训练图像S1和S2，通过分析水稻S1和背景S2两类图像在RGB色彩空间中R、G、B色彩特征值的分布频率，rgb（标准化RGB）色彩空间中r、g、b色彩特征值的分布频率，CIE L*a*b*色彩空间中L*、a*、b*色彩特征值的分布频率，以及HSV色彩空间中H、S、V色彩特征值的分布频率，筛选出具有明显双峰特征的g、a*、b*和S作为关键色彩特征；然后，在由g、a*、b*和S色彩特征构成的多维色彩空间中采用支持向量机学习算法获得分隔水稻和背景像元的优化超平面Z（Z=0.421g+0.753a*+0.152b*+0.051S+0.085）；最后，计算水稻冠层图像中每一像元的Z值，并用最大类间方差法计算分割阈值Zt，从而实现水稻冠层图像分割。为了评价该分割方法，以90幅不同光强（阴天、多云和晴天）条件下采集到的田间水稻冠层图像作为测试图像，并以常用的ExG & OTSU分割方法作为对比，分析该分割方法的分割精度和光强稳健性。结果表明，ExG & OTSU方法的精确度显著受光强条件影响，随着光强强度的提高而显著降低，光强稳健性差；该研究所提分割方法对阴天、多云和晴天条件下水稻冠层图像的分割误差为7.30%、8.72%和8.98%，分割精确度较高，且具有良好的光强稳健性。因此，该基于支持向量机和最大类间方差法相结合的水稻冠层图像分割方法具有较高的分割精度和光强稳健性，可为田间多变光照条件下水稻冠层图像精准分割提供技术参考。

关键词: 支持向量机, 最大类间方差法, 色彩空间, 水稻, 光照强度, 图像分割