基于概率局部判断映射的植物分类方法

基于叶片图像的植物分类与识别方法研究在保护植物物种和生态环境等方面发挥着重要的作用。由于叶片图像的复杂多样性以及同类叶片图像之间的差异性较大等特点,使得很多基于叶片颜色、形状和纹理的特征提取和识别方法不能满足植物自动识别系统的需要。在分类概率和局部保持映射(locality preserving projections,LPP)的基础上,提出了一种概率局部判别映射(probability locality preserving discriminant projections,PLPDP)方法,并应用于植物分类。首先计算每个样本的分类概率,由样本的局部信息、分类概率和类别信息定义权重矩阵,然后构建目标函数。通过最小化目标函数寻求最佳投影矩阵,使得原始高维样本经投影后,在低维特征空间保持了样本的局部信息、分布信息和类别信息。与判别LPP和监督LPP相比,PLPDP充分利用了样本的局部信息、分类概率和类别信息,算法的分类能力得到了较大提高。在公开的植物叶片图像数据库上对20类植物叶片图像进行了分类试验,识别率高达90%以上。试验结果表明,该方法是有效可行的。

Probability locality preserving discriminant projections for plant recognition

Abstract: Study on the classification and recognition methods of plant species by using plant leaf images plays an important role in protecting plant species and ecological environment. Designing a computer-aided plant recognition system is necessary and useful since it can facilitate plants recognition and classification, and understanding and managing plant species. Compared with other plant recognition methods, such as cell and molecule biology methods, plant recognition and classification based on leaf image processing is becoming a popular trend. In protecting plant perspective, leaf images have been used by plant protection researchers to diagnose plant diseases and this method has been proven to be reliable for years. Each kind of plant leaf has its own features and carries large significant information that can be used to recognize and classify the origin or the type of plant. Leaf shape is a prominent feature that most people use to recognize and classify a plant. The features, such as leaf area, perimeter, diameter, physiological length and physiological width, are basic geometry information that can be extracted from the leaf shape. In addition, leaf color, textures and vein pattern are also considered as important classifying features. All these features are useful for recognizing and classifying plant. Because of the complex and diversity of plant leaf images and the differences between within-class leaf images, many classification and recognition methods that use color, shape and texture of the leaves cannot meet the need of the plant automatic identification system. The feature extraction and dimensional reduction is a key step to plant classification. The classical linear dimensional reduction methods can not effectively applied to leaf image processing because the plant leaf images are general nonlinear data. Manifold learning based recognition methods have been successfully applied to face recognition. Based on manifold learning, a probability locality preserving discriminant projections (PLPDP) method was proposed for plant recognition in this study. The method was to learn a linear transformation, and focuses only on the pairwise points where the two points were neighbors to each other. First, the classification probability of each sample was computed, and the weighted matrix was defined by local information, label information and classifying probability. Then objective function was redefined. Then, through minimizing the function, an optimal projection matrix could be preserved in the low-dimensional feature space, such as the distribution information contained in the original data. Compared with the popular dimensionality reduction methods, such as linear discriminant analysis (LDA), locality preserving projections (LPP), discriminant LPP (DLPP) and supervised LPP(SLPP), the proposed method took three kinds of information into account (i.e., local information, label information and classifying probability) and refined the weighted matrix, which not only contained neighborhood information of samples, but also could reflect the probability from which a sample was correctly classified when its K-nearest-neighbors were selected. After projection, the neighborhood relationship of the intra-class samples which possessed more classification probability could be preserved, and the considered pairwise points in within-class were as close as possible, while those in between different classes were as far as possible, in which the global optimum could be effectively obtained. Finally, the experimental results on the public plant leaf image database demonstrate the superiority of the proposed algorithm compared with other algorithms.