【作者】 王晓松；

【导师】 黄心渊；

【作者基本信息】 北京林业大学 ， 林业装备工程， 2010， 博士

【摘要】 随着计算机信息技术的不断发展,“精准林业”思想提出以后,林业立体视觉测量、林木农药精确对靶施用、基于图像的树木可视化重建、生长状态评估、树种自动识别与分类等课题被林业科研工作者提出并不断研究探索。树木图像提取为以上研究提供基础数据和技术支撑,是以上研究的一个重要基础和难点问题。树木本身及其周围景物的多样性使得复杂背景中的树木图像提取成为一项复杂的、探索性很强的工作。因此,复杂背景下树木图像提取方法和技术的研究具有重要的实用价值和现实意义。本文以树种的机器识别为研究背景,针对复杂背景下树木图像的特有特征,进行复杂背景下树木图像提取的分割技术和自然图像抠图技术研究。图像分割基础方法在树木图像提取中的应用研究。图像分割方法主要可分为阈值分割、边缘检测、区域分割三种基本类型。本文用基础的图像分割方法对复杂背景下的树木图像进行分割,根据提取结果分析了三种基础的图像分割方法在复杂背景下树木图像提取中的局限性。基于颜色特征和纹理特征的树木图像分割方法研究。色彩的特征是树木有别于环境的最大特征,有利于分离图像中的非绿色植物和非绿色背景。树木与背景中的其他绿色植物的纹理也存在明显差别,所以也可以利用纹理的统计分析有效的分割树木图像。本文提出了一种结合颜色和纹理特征的树木图像分割设计,把图像从RGB空间转化到Lab色彩空间,然后把a通道分离出来,再根据灰度共生矩阵算法提取图像的纹理特征,最后对灰度图像的颜色和纹理特征进行分割以及数学形态修正。抠图技术提取复杂背景下树木图像研究。一幅自然场景中的树木图像的树叶边缘通常比一个像素还要细小,或者目标树木周围包含与目标树木非常相近的绿色植物等。论文对现有的几种自然图像抠图方法进行分析比较,总结了各种方法的特点及其在树木图像提取中的局限性。实现了GrabCut抠图提取树木图像的方法。该方法需要少量的人工交互,对于前景和背景颜色区别明显的树木图像分割效果好。改进的基于马尔科夫随机场的树木图像抠图技术研究。树木图像具有前景空洞很多、伴有透明和半透明现象的特征。本文提出了关注区域的概念,实现了背景区域的自动标示功能,并把区域生长与抠图技术相结合。从简化三分图划分、尽可能多地确定前景像素点和减少未知区域像素数目三个角度对基于MRF的抠图方法进行改进,得到了一种快速、有效,而且实用的树木图像抠图方法。图像分割方法与自然抠图技术提取树木图像的比较研究。本文对综合纹理与颜色特征的树木图像分割方法和改进的MRF抠图进行了系统分析、设计与实现。对多幅原始图像应用图像分割和自然图像抠图两种方法进行提取,并对提取成功率、提取速度和人机交互量进行统计分析。两者差异表现为：图像分割方法算法相对简单,运算速度较快,无需人工交互,但是抠图成功率较低；自然图像抠图方法则算法相对复杂,运算速度较慢,需要适量的人工交互,但是抠图成功率高。更多还原

【Abstract】 Tree image extraction is a technology to separate an object tree from the surrounding landscape in a photograph which is shot on the ground, and to obtain its characteristics data. With the continuous development of computer information technology, the thinking of Precision Forestry is proposed. Forestry researchers made related topics such as forestry stereo vision measurement, accurate to the target application of pesticides tree, image-based visual reconstruction of trees, growth, condition assessment, automatic identification and classification of tree species. And they will continue to study and explore these fields. Tree image extraction provide the basis data and technical support for the above research and applications, it is very important and still a difficulty problem. The background of a photograph which is shot in the natural scene is uncertain and complex. The diversity of trees and surrounding scenery make the work to extract an object tree in the complex background is hard and highly groping. This research has important applied value and practical significance. This paper take the machine identified as a research background, for the purposes of improve the speed, quality of matting and further reduce the human interaction involvement. Based on existing technology, in the image segmentation and natural image matting for complex background image of the unique characteristics of trees, make research and implementation of tree appearance feature extraction method for complex background image of the unique characteristics of trees.Firstly, the paper researches the image segmentation method to extract the image of trees in the applied. Image segmentation method can be divided into threshold segmentation, edge detection, region segmentation. With the continuous development of imaging technology, some segmentation methods are proposed such as image segmentation based on specific theory, combination of multi-segmentation method and human-computer interaction image segmentation. This paper analysis the shortcomings, limited and improvements according to the segmentation results of the current main method. Then, according to the main features of the trees, this paper carried out image segmentation based on color features and texture features. The characteristics of color is greatest different between the trees and the environment. The use of color features is conducive to separation of non-green plants and non-green background. Canopy of different tree species often have significant texture feature differences, also trees and other green plants in the background have significant texture difference. Therefore, statistical analysis of texture can be used to effectively split the images of trees. In this paper, first transform the color space of images from RGB to Lab, then separate a channel, extract image texture feature according to the Gray Level Co-occurrence Matrix algorithm, thought segment gray image of color and texture features and apply mathematical morphological amendment. This method is relatively simple, fast and has markedly accuracy compared to traditional methods. However, the method could do nothing if the background is too complicated.On top of the research to extract tree images by means of image segmentation, there has studied the application, of natural image matting technology in tree image extraction. In an image of trees, the edge of a leaf in nature scene is usually even smaller than a pixel, or there are green plants around very similar to the goal objectives, etc. In these complex cases, image segmentation method cann’t work efficient. There summed up the limitations of several existing extracting methods through analysis and comparison, and realized GrabCut algorithm.Based on the above analysis, this paper presents an improved natural image matting techniques based Markov Random Field. There would be a lot empty in foreground of tree image, also accompanied by the phenomenon of transparent and translucent features. Taking into account this situation, this paper improved MRF method by simplified one-third figure, broken down as much as possible to determine the prospects for the unknown pixel and reduce the number of three point of view of regional pixel-based. This method reduces the complexity of human-computer interaction, markedly enhanced the color accuracy. Also enhance the computing speed more than 34%, it is a very efficient and practical trees image matting method.Finally, the paper compares and makes research of the image segmentation method and the natural image matting method for the tree image extraction technology according to the results. By statistical analysis, the differences of two methods were expressed as:image segmentation method is relatively simple and relatively fast operation, without human interaction, but the relatively low accuracy rate; natural image matting methods are complexity, computational speed slower, need the right amount of human interaction, but the extraction rate is high accuracy.更多还原