不同光照条件下农田图像分割方法的研究

发布时间：2018-05-12 21:12

本文选题：农田图像分割 + 颜色因子　；参考：《西北农林科技大学》2017年硕士论文

【摘要】：由于受到天气、温度和光照等因素的影响,智能农业机器人感知环境信息时会存在一定的不确定性。为进一步提高智能农业机器人的环境感知能力,需对不同光照条件下的农田图像进行分割。本研究以西北农林科技大学北校区试验田三叶期至五叶期玉米农田图像为研究对象,采用颜色因子法,结合阈值分割法和机器学习法实现了不同光照条件下的农田图像分割,并通过主观评价法和客观评价法完成算法分析及验证。本研究的主要内容和结论有:(1)农田图像的获取及分类。为实现获取农田图像的自动化分类,提出了基于数学统计学知识分析农田图像直方图的方法。实验发现,不同光照条件下农田图像R,G,B颜色通道对应的直方图,其均值指标和偏度指标在任何区间上均没有重合,可作为农田图像自动分类的标准。与人工分类方法对比后发现,本文方法分类误差率最大为10.52%,说明采用上述方法可实现农田图像的自动分类。(2)光照充足或光照偏弱条件下的农田图像分割。针对光照充足条件下农田图像颜色特征较为明显的特点,主要采用直方图均值法和大津法实现了农田图像的分割;针对光照偏弱导致农田图像颜色和形状特征不显著的特性,主要采用无监督学习中的模糊C均值聚类算法(FCM)实现了农田图像的分割。最后完成两类实验结果的剖析比较。由于光照充足和光照偏弱条件下农田图像分割目标十分复杂,因此主要采用了主观评价法分析实验结果。实验发现,两类农田图像分割结果平均主观质量分数分别为4.26和4.06,则根据CCIR500五级评分质量尺度和妨碍尺度说明,采用本文方法图像分割质量较好,可实现复杂农田图像的分割。(3)光照偏强条件下的农田图像分割。针对光照偏强导致大量高光点对图像分割精度干扰的问题,本文提出采用改进的简单线性迭代聚类算法(SLIC)完成图像预处理提取超像素,提取特征向量并通过曲线进行初步筛选,然后建立分类器实现农田图像的分类。分类器主要选择贝叶斯和支持向量机(SVM)。实验发现,改进的SLIC在不影响图像预处理结果的前提下可缩短运行时间;SVM总体分类精度优于贝叶斯,平均总体分类精度可达到94.83%,说明采用SVM可有效实现含大量高光点简单农田图像分割。以农田图像自动分类为研究基础,本文基本完成了不同光照条件下的农田图像分割,为提高智能农业机器人感知环境信息能力提供了有力的保障。
[Abstract]:Due to the influence of weather, temperature and light, the intelligent agricultural robot will have some uncertainty when it perceives environmental information. In order to improve the environment perception ability of intelligent agricultural robot, it is necessary to segment farmland images under different illumination conditions. In this study, the field images of maize in three leaf period to five leaf stage in the experimental field of North Campus of Northwestern University of Agriculture and Forestry Science and Technology were studied. Using color factor method, combining threshold segmentation method and machine learning method, the field image segmentation under different illumination conditions was realized. The algorithm is analyzed and verified by subjective evaluation and objective evaluation. The main contents and conclusions of this study are: 1) farmland image acquisition and classification. In order to achieve automatic classification of farmland images, a histogram analysis method based on mathematical statistics was proposed. It was found that the histogram corresponding to the color channel of RDG _ (B) in farmland images under different illumination conditions had no coincidence in any interval, so it could be used as a standard for automatic classification of farmland images. Comparing with the artificial classification method, it is found that the maximum classification error rate of this method is 10.52, which shows that the above method can be used to realize the automatic classification of farmland images with sufficient illumination or weak illumination. In view of the obvious color characteristics of farmland images under sufficient illumination, the histogram mean method and Otsu method are mainly used to segment farmland images, and the weak illumination leads to the characteristics that the color and shape features of farmland images are not significant. The fuzzy C-means clustering algorithm (FCM) in unsupervised learning is used to segment farmland images. Finally, two kinds of experimental results are analyzed and compared. Because the target of farmland image segmentation is very complex under the condition of sufficient illumination and weak illumination, the subjective evaluation method is mainly used to analyze the experimental results. The experimental results show that the average subjective mass scores of the two kinds of farmland images are 4.26 and 4.06, respectively. According to the quality scale and hindrance scale of CCIR500 five-grade score, the image segmentation quality is better by using this method. The segmentation of complex farmland image can be realized under the condition of strong illumination. Aiming at the problem that a large number of high light points interfere with image segmentation accuracy due to the strong illumination, an improved simple linear iterative clustering algorithm (SLICs) is proposed to extract super-pixels, extract feature vectors and screen through curves. Then a classifier is established to realize the classification of farmland images. The classifier mainly selects Bayes and support Vector Machine (SVM). The experimental results show that the improved SLIC can shorten the running time and the overall classification accuracy is better than that of Bayes without affecting the result of image preprocessing. The average overall classification accuracy can reach 94.83, which shows that using SVM can effectively realize the segmentation of simple farmland images with a large number of high light points. Based on the automatic classification of farmland images, the segmentation of farmland images under different illumination conditions is basically completed in this paper, which provides a powerful guarantee for improving the ability of intelligent agricultural robots to perceive environmental information.
【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S126;TP391.41

【参考文献】