基于颜色特征的绿色作物图像分割算法

摘要

绿色作物的识别是农业机械视觉系统的重要研究内容之一,该文采用RGB颜色系统,基于统计分析提出了一种绿色作物图像分割方法.从简单物体光照颜色模型方面,分析了RGB颜色空间中作物绿色“恒量”(Gvalue＞Rvalue and Gvalue＞Bvalue)的存在性,构建了作物图像分割相对错误率评估模型.并与传统颜色索引方法Excess Green(ExG)+auto-threshold进行了对比分析.试验结果表明,在正常光照条件下:1)采用的算法对田间不同作物-土壤组图像分割的相对错误率均有显著影响:其中,相对ExG+auto-threshold算法,采用RGB算法的结果图像中大多能保留油菜、大豆和甘蔗的形态学特征；2)采用的算法、光照变化以及算法与光照变化的交互作用均对室外美人蕉图像分割的相对错误率有显著影响；其中,相对ExG+auto-threshold算法,采用RGB算法的结果图像中大多能去除背景噪声.单因子方差分析进一步表明,光照变化对采用ExG+auto-threshold算法分割图像的阈值有显著影响.该文提出的RGB算法相对传统的ExG+auto-threshold绿色索引,对于早期生长的绿色作物是一种有效、简单的图像分割方法,对作物-土壤、光照变化不敏感.%Crop rows recognition is a principal issue in agricultural machinery vision system. In this paper, the color constant in RGB space was found to segment the green plant from the background by using statistical analysis based on the classical simple illumination model. For early green vegetation, the value of green component Gvalue is always greater than that of the other red Rvalue and blue 5value components (Gvalue>Rvalue and Gvalue >Bvalue, inferred to as RGB); and a relative segmenting error ratio was designed to evaluate the performances of the RGB presented in this paper and the Excess green + auto-threshold(ExG+atuo-threshold). In Experiment 1, the single factor variance analysis showed the algorithms (RGB and ExG+auto-threshold) had significant effect on the relative segmentation error ratio of the plant-soil images, and the corresponding segmented image using RGB were found that most of them could preserve morphological feature of plant compared with ExG+auto-threshold. And in Experiment 2, the double factor variance analysis showed that the algorithms, illuminant variations and their interaction had significant effect on the relative segmentation error ratio of the Canna images grasped consecutively, and the illuminant variations affected the threshold values of ExG+k-auto. The corresponding segmented images using RGB were found that most of them could delete the background noises compared with ExG+auto-threshold. And therefore, the RGB is a simple but efficient segmentation algorithm, and insensible to plant-soil and illuminant variations compared with ExG+auto-threshold.