依据特征融合和深度学习的树木叶片分类方法

Classification Methods of Tree Leaves by Feature Fusion and Deep Learning

以Pl@ntNet Identify、leafsnap等树木叶片数据库中的9500张图片为样本,将叶片的特征融合后作为分类依据;将改进的局部三值模式特征和梯度方向直方图特征采用零均值标准化方法进行融合,采用深度信念网络进行训练、识别和分类。结果表明:融合测试方法识别率可达94.87%,优于其他方法在本数据库的识别率;融合方法比单一特征和支持向量机分类等方法识别率更高,且受光照、噪声等影响的鲁棒性更高;实现了树木叶片的快速识别,解决了依据特征的叶片分类方法识别率较低的问题,改善了已有方法特征选取单一、信息不足和分类器简易等不足。

We take 9500 pictures in the tree leaf database of Pl@ntNet Identify,leafsnap and other trees as samples,and used the features of the leaves as the classification basis,the improved local ternary pattern feature and histogram of oriented gradients feature are fused by using the zero mean normalization method.We used deep belief networks for training,recognition and classification.The recognition rate of the fusion test method can reach 94.87%,which is better than the recognition rates of other methods in this database.The fusion method has a higher recognition rate than methods such as single feature and support vector machine classification.The robustness affected by light and noise is higher.The fast recognition of tree leaves is achieved,the problem of low recognition rate of leaf classification methods based on features is solved,and the existing methods have the disadvantages of a single feature selection,insufficient information,and simple classifiers.