基于Lab颜色空间的非监督GMM水稻无人机图像分割

Unsupervised GMM for Rice Segmentation with UAV Images Based on Lab Color Space

摘要:

为解决传统水稻冠层图像分割算法性能在很大程度上依赖于训练数据集的质量，且分割效果易受田间多变光照强度影响，导致水稻生产参数估计精度不高等问题，提出一种基于Lab颜色空间的非监督贝叶斯方法，用于田间水稻无人机图像分割。模型参数从每个独立、未标记的无人机图像直接学习获得，无需训练。不同图像会有不同的模型参数，该算法能够适应各种不同环境拍摄的图像。将提出的算法应用于分蘖后期田间水稻的无人机图像分割，并与RGB-GMM、HSV-GMM和All-GMM算法进行对比，在高度10m图像中平均查全率、平均查准率和平均F1值分别为0.8427、0.7570和0.7948，在高度15m图像中分别为0.8756、0.7133和0.7788，优于RGB-GMM、HSV-GMM和All-GMM算法。研究表明，本文提出的方法可以从复杂大田环境拍摄的无人机影像中准确提取水稻像素。

Abstract:

Rice image segmentation is a key step to obtain rice growth parameters, and plays an important role in rice production. The performance of traditional rice canopy image segmentation algorithm largely depends on the quality of the training data set, and the segmentation result is easily affected by the variable light intensity in the field, which leads to the poor estimation accuracy of rice growth information. In order to solve the above problems, an unsupervised Bayesian method based on Lab color space was proposed for field UAV image segmentation. With the unsupervised learning approach, the model parameters were directly learned by using unlabeled data from each individual UAV image. Different images had different model parameters, and this made the algorithm adaptable to images taken under a wide variety of conditions. The proposed algorithm was applied to UAV image segmentation of rice field in late tillering stage, and compared with RGB-GMM, HSV-GMM and All-GMM algorithms. Applying the algorithm on diverse UAV images in 10m height achieved an average recall, precision and F1 score of 0.8427, 0.7570 and 0.7948, respectively. Applying the algorithm on diverse UAV images in 15m height achieved an average recall, precision and F1 score of 0.8756, 0.7133 and 0.7788, respectively. These numbers outperformed the RGB-GMM, HSV-GMM and All-GMM algorithms. The experimental result demonstrated that the proposed method can accurately identify rice pixels in UAV images taken under diverse conditions.