基于多特征融合和深度信念网络的植物叶片识别

Plant leaf identification based on the multi-feature fusion and deep belief networks method

基于叶片数字图像的植物识别是自动植物分类研究的热点。但是随着植物种类的增加,传统的分类方法由于提取的特征比较单一或者分类器结构过于简单,导致叶片识别率较低。为此,本文提出使用纹理特征结合形状特征进行识别,并且使用深度信念网络构架作为分类器。纹理特征通过局部二值模式、Gabor滤波和灰度共生矩阵方法得到。而形状特征向量由Hu氏不变量和傅里叶描述子组成。为了避免过拟合现象,使用"dropout"方法训练深度信念网络。这种基于多特征融合的深度信念网络的植物识别方法,在Flavia数据库中,对32种叶片的识别率为99.37%;在ICL数据库中,对220种叶片的识别率为93.939%。这表明相比一般的叶片识别方法,此方法鲁棒性更强,并且识别率更高。

Plant identification based on digital leaf image is a hot topic in research of automatic classification of plants. However,with the increase in the number of plant species,the leaf recognition rate is low due to the single trait extraction and simple structure classifier in the traditional classification methods. This study applied the combination of texture features and shape features for identification,using the deep belief networks（ DBNs） as the classifier. Texture features are derived from local binary patterns,Gabor filters and gray level co-occurrence matrix while shape feature vector is modeled using Hu Moment invariants and Fourier descriptors. In order to avoid overfitting,we trained the DBNs with ＂dropout ＂method. The proposed algorithm was tested on the Flavia dataset,and the recognition rate was 99. 37%for 32 species,while on the ICL dataset the recognition rate was 93. 939% for 220 kinds of leaves. The experimental results illustrated that the proposed method has stronger robustness and higher recognition rate compared to the traditional methods.