结合LDA主题模型的植物叶片形状描述及分类

Leaf Shape Description and Classification with LDA Topic Mode

针对传统形状描述算子多侧重于叶片形状再表达,对叶片形状随机特性刻画不足的缺点,结合潜在狄利克雷分布(LDA)主题模型建立有效的植物叶片形状描述算子,并据此构建叶片形状识别分类框架.首先建立叶片形状的多尺度词包模型,将形状空间联系引入形状生成模型.然后结合LDA建立叶片形状生成模型,提取形状分布参数作为叶片形状描述算子.最后使用K近邻进行叶片分类.实验表明,在异类叶片类间形状差异较小的复杂情况下,相比傅里叶、形状上下文等传统算子,结合LDA主题模型的植物叶片形状描述算子的叶片形状识别精度更高.

Since the conventional shape descriptors focus on shape reexpression and the stochastic character of leaf shape is neglected, an effective plant leaf shape descriptor is proposed based on latent Dirichlet allocation （LDA） model. A corresponding leaf shape recognition framework is also constructed. Firstly, the plant leaf shape is transformed and represented as a multi-scale bag-of-words model, and thus the space interaction relationship is introduced into the leaf shape generative model. Furthermore, a leaf shape generative model is established via LDA model, and then the leaf shape descriptor is designed by the extracted shape distribution parameters in the LDA model. Finally, k-nearest neighbor （KNN） method is applied to the plant leaf shape classification. Experimental results demonstrate that the leaf shape descriptor combined with LDA model effectively improves the shape classification accuracy, especially for the plants of different classes hut with a roughly similar shape of leaf. The proposed method obtains a higher classification accuracy compared with the conventional shape descriptor.