中文摘要：

针对基于叶片特征进行树种识别的问题,本文在结合叶片纹理、不变矩以及传统形状共25维传统特征的基础上,自定义了叶尖角、边角均值等2个叶片轮廓特征,并以相似多边形定义及其推论作为理论依据,提出了一种基于叶片轮廓构建距离矩阵与角点矩阵进行树种识别的分类方法。该方法首先对树木叶片图像进行预处理,提取出归一化的叶片特征向量,然后利用KNN最近邻分类器筛选出相似度最高的前20个结果集（Top 20）,然后构建距离矩阵和角点矩阵进行更为精确的识别匹配。在图像预处理阶段,为获取更为准确的叶片轮廓特征,利用叶片在HSV颜色空间中饱和度特征以及色度特征方面的显著差异性,设计了一种消除叶片阴影的图像预处理算法。在识别匹配阶段,利用Douglas Peucker approximation算法提取叶片轮廓的近似多边形,定义了距离矩阵、角点矩阵、矩阵中元素间相似度、矩阵相似度以及综合相似度计算方法,设计了全局匹配与局部匹配相结合的算法。该算法在Android系统的手机平台上进行了实现和运行验证,结果表明：在Flavia数据集中,对32种共1 907个正常叶片样本的识别准确率为99.61%,对32种共851个残叶样本的准确率为94.92%;在Leafsnap数据集中,对185种共23 147个Lab样本前5个结果集（Top 5）的识别准确率为98.26%。相对其他算法,该算法识别准确率更高,对叶片外形描述能力更强,对残叶、扭曲叶、阴影叶具有更好的鲁棒性,算法的实用性和适应性更强。

英文摘要：

Aiming at the problem of tree species identification based on leaf characteristics, this paper self- defines two leaf outline characteristics of leaf angle and edge angle mean on the basis of 25 kinds of characteristics such as leaf texture, invariant moments and traditional shapes, regards the definition and inference of similar polygon as theoretical basis and proposes a kind of classification method for tree species identification which is to construct distance matrix and corner matrix based on leaf outline. Firstly, the tree leaf image was pre-processed to extract the normalized leaf feature vector. Then KNN was used to choose the top 20 of the highest similarity result set. And then distance matrix and corner matrix were constructed to identify and match more accurately. During pre-processing, in order to obtain more accurate characteristics of leaf outline, this paper designs an image preprocessing algorithm of eliminating leaf shadow by taking advantage of significant differences between saturation and chromaticity in HSV color space. During identification and matching, this paper uses the Douglas Peucker approximation algorithm to extract the approximate polygon of leaf outline, and defines calculation methods of distance matrix, corner matrix, matrix element similarity, matrix similarity and comprehensive similarity, and designs a kind of algorithm combining global matching and local matching. The algorithm had beenimplemented and run on Android mobile phone platform. Results showed that the accuracy of the algorithm was 99.61% among 1 907 complete samples of 32 categories and 94.92% among 851 incomplete samples of 32 categories in Flavia dataset. In Leafsnap dataset, the accuracy of the algorithm was 98.26% of top 5 among 23 147 Lab samples of 185 categories. Compared with other algorithms, this kind of algorithm had higher identification accuracy, better description ability for leaf outline, better robust performance for incomplete leaf, twisted leaf and shadowed leaf, and better practicability and ada