基于随机森林的玉米发育程度自动测量方法

为提高玉米果穗发育程度检测的自动化程度与精度,提出一种基于机器视觉技术的测量方法。在随机森林机器学习算法的基础上构造秃尖、干瘪和籽粒区域的识别模型。该模型由多个独立同分布的弱分类器构成,对输入的训练样本进行列和行两个方向上的随机采样。比较随机森林模型和决策树模型的分类效果可知随机森林模型有效避免了过拟合和局部收敛现象的产生,并具有良好的推广能力。为确定最优的弱分类器数目,选择弱分类器个数为训练样本数量的1/80、1/40、1/20、1/10、1/5、1/4时分别构建随机森林分类器。研究结果表明,当随机森林中弱分类器个数为训练样本数量的1/20时,模型的识别率与稳定性最好。然后,以最优的随机森林模型作为分类器构建玉米果穗不同发育程度自动检测方法。试验结果表明,各区域长度测量的准确性均在95%以上,测量速度可达30个/min以上。

Automatic Measurement Method for Maize Ear Development Degree Based on Random Forest

In the process of maize breeding, the development degree of maize ear is one of the most important parameters for yield related traits. In order to improve the degree of automation and accuracy of maize ear development degree detection, a measurement method was proposed based on machine vision technology. An identification model was constructed on the basis of random forest principal at first. The model was composed of a group of weak classifiers which were independent and identically distributed. The weak classifiers selected samples from the input training samples randomly along columns and rows. The experiment which compared random forest model with decision tree model on the classification effect showed that random forest classifier could not only avoid over-fitting and local convergence effectively but also have good generalization ability. Then, in order to determine the optimal number of weak classifiers, six random forest models were built. Their weak classifier number were separately one-eightieth, one-fortieth, one-twentieth, one-tenth, one-fifth, one-fourth of training samples count. The results showed that the model had good accuracy and stability when the number of weak classifiers was one-twentieth of training samples count. Finally, the optimal random forest model was used as the classifier to build the automatic maize ear development degree detection method. The experiment results showed that the measurement accuracy on length of each area was more than 95% and the measurement speed was more than 30 maize ears per minute.