一种基于K-means与Close-Form融合的树木图像提取方法

A Tree Image Extraction Method Based on K-means and Close-Form

基于图像的树木自动提取是其种类、长势、形态等信息智能化判别的基础,如何实现树木的自动、准确、快速提取是具有实用性的科学问题。在自然场景中,由于图像元素多样、颜色差异大,树木自身存在不规则性,树木提取难度非常大。针对现有的图像分割与图像抠图法在树木提取过程中分别存在的误分割与过程复杂所导致的计算量大的问题,提出了一种基于K-means聚类算法优化Close-Form图像抠图的树木提取方法。在少量的标记下,依据颜色线性假设进行最小化代价函数计算,得到图像透明度;对透明度图像依次进行中值滤波、高斯滤波,得到透明度的去噪图像;对滤波后透明度与标准化的图像绿色分量组成二维空间进行K-means二聚类,实现背景与前景对象的准确判定,进而完成自然背景的树木图像提取。为了验证所提方法在不同场景和不同标记下的树木提取有效性,设计了基于K-means图像分割和传统Close-Form抠图方法的比较性试验。结果表明,基于K-means优化Close-Form的树木提取方法解决了传统Close-Form算法在少量标记下图像前景、背景估计不准确问题,克服了图像分割存在的误分类情况,实现了不同自然环境和多目标树木对象的提取。此方法具有对象提取稳定、计算时间快的优点,相对原Close-Form算法用时减少49.98%。

Image-based automatic extraction of trees is the basis for the intelligent identification of species,growth,morphology and other information.How to realize the automatic,accurate and rapid extraction of trees is a practical scientific problem.In the natural scene,due to the diversity of the image elements,the large differences in color and the irregularity of the trees themselves,the extraction of trees is very difficult.In this paper,aiming at the problem of large amount of computation due to the error segmentation and complicated process in the existing tree image segmentation and image matting methods,a K-means clustering algorithm was proposed to optimize Close-Form images cutout tree extraction method.Firstly,the transparency of the image was calculated by minimizing the cost function under the linear color hypothesis with a small number of labels.Then,the transparency image was sequentially filtered through median and Gaussian filters to obtain the transparency of the denoised image.Finally,K-means clustering of filtered transparency and image green components was conducted to achieve the accurate determination of the background and foreground objects,and then the natural background of the tree image extraction was accomplished.In order to verify the effectiveness of tree extraction in different scenes and different markers,a comparative experiment based on K-means image segmentation and traditional Close-Form matting method was designed.The experimental results showed that the optimization of Close-Form method of the tree extraction solved the problem of inaccurate image foreground and background estimation of the traditional Close-Form algorithm with few marks and overcame the misclassification of image segmentation.The method could achieve the extraction of objects in different natural environments and multi-target trees.In addition,the method had the advantages of stable object extraction and fast calculation time,compared to the original Close-Form algorithm,the time was reduced by 49.98%.