基于Graph Cut算法的多株立木轮廓提取方法

【目的】在复杂的自然环境下进行目标立木轮廓提取时,容易受遮挡物影响,导致立木图像分割效果不理想。笔者提出一种基于Graph Cut算法的多株立木轮廓提取方法,可实现单张相片中多目标立木界线分割。【方法】首先通过彩色直方图均衡化实现RGB颜色空间下各个通道的图像细节增强,利用Graph Cut算法构造s-t网络图,将图像分割问题转化为能量函数最小化问题,并标记图像前背景像素实现单张相片中多株立木图像初分割; 然后将单张相片中的每株立木分割图像二值化,利用形态学腐蚀膨胀运算处理图像达到填充、去噪、平滑等目的; 在此基础上,利用改进型Canny算子边缘检测方法,用双边滤波代替高斯滤波增强边界信息得到每株立木轮廓; 最后,根据立木相对坐标不变性,利用几何重组方法实现目标立木特征表达并判断其拓扑关系,最终得到每株目标立木轮廓提取结果。【结果】为了验证该方法的有效性,本研究对自然环境下采集到的立木图像进行试验。结果表明,该方法能够在不同光照条件的复杂背景下,有效分割出每株立木轮廓,平均误分率为5.62%,假阳性率为4.49%,假阴性率为4.33%,均优于常用的OTSU分割算法(41.40%、26.73%、10.99%)、K-means聚类算法(49.97%、35.02%、11.92%)和基于C-V模型水平集法(28.43%、20.53%、13.38%)。【结论】复杂的自然环境下,利用基于人工交互的Graph Cut算法可有效分割出每株立木轮廓界,研究结果可为立木可视化重建、特征提取等提供参考。
【Objective】 Because of the complexity of the natural environment, current tree contour extraction results are not satisfactory. This paper presents a method to extract a contour of multiple trees based on a Graph Cut algorithm to realize the boundary segmentation of multi-target trees in a single photo. 【Method】 First, this method enhances image details of each channel under RGB color space captured in the experiment by color histogram equalization. The graph of the s-t network is constructed using a Graph Cut algorithm to look for min-cut, and the image segmentation problem is transformed into the minimization of the energy function by marking the foreground and background pixels to achieve a single photo of many trees with preliminary image segmentation. Then, adaptive thresholding of gray-scale transformation is applied to the multiple-segmentation images to realize binarization of the images and morphological corrosion expansion, and the opening and closing operation processing of the binary images is used to achieve the filling, denoising, and smoothing of trees. On the basis of this morphological process, combined with the improved Canny operator edge detection technology, bilateral filtering is used instead of Gaussian filtering to enhance the boundary information to obtain a preliminary tree contour. Finally, according to the geometric position invariance of the photo’s trees, we use the geometric reconstruction method to express the features of target trees and judge their topological relationships. If there are topological relationship errors, we iterate the Graph Cut algorithm and geometry reassembly method again to obtain a better target tree extraction result. 【Result】 In order to validate the effectiveness of this method experimentally, we collected tree images in a natural environment. The results showed that this method can effectively separate the contour of every tree under different