基于复合叶片特征的计算机植物识别方法  

Plant Recognition Based on Compound Leaf Features

作　　者：高翔[1] 王正[1] 丁见亚 杨倩[1] 

机构地区：[1]清华大学自动化系,北京100084

出　　处：《植物学报》2014年第4期450-461,共12页Chinese Bulletin of Botany

摘　　要：该文探讨如何根据植物的叶片特征,利用图像处理和机器学习的方法对植物进行分类。鉴于现有的叶片分类系统多采用单一的特征,如几何和纹理等,仅能在小规模数据库上得到较好的结果。然而,随着样本种类的增多,单一特征在不同种类叶片之间的相似性非常明显,致使分类正确率降低。该研究使用多种复合特征,并提出了原创的预处理方法以及宽度、叶缘频率特征,较传统的几何特征更为详尽。研究结果显示,复合特征可以有效避免算法过拟合问题,使之适用于更大的数据库。通过提取21类植物的叶片宽度、颜色、叶缘和纹理共292维特征,对1 915张数字图像进行了分类,正确率达到93%,并分析了各类特征对分类结果的影响。研究结果表明,在不影响分类正确率前提下,可将特征减少到约100维。This study investigated how to classify plant species with compound leaf features using machine-learning approaches. Many traditional classification systems used a single feature, such as geometry or texture. Although such systems can achieve good results in small databases, with increasing records, the similarity in single features between different species will be remarkable, thus reducing the accuracy in large databases. This study examined how to extract compound features and proposes a novel preprocessing method and new ways to extract width and edge information, which are more detailed than with most of the state-of-the-art approaches. The compound features can reduce the influence of the over-fitting problem, so the algorithm can be used for larger databases. We examined up to 21 kinds of plants （extracting width, color, edge and texture data） and 1 915 digital images and achieved an accuracy of 93%. Finally, we analyzed the effect of each feature on classification results. We could reduce the feature＇s dimension to about 100 without losing much classification accuracy.

关 键 词：机器学习 模式识别 植物分类 图像处理 叶片特征 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]