基于双路卷积神经网络的植物叶片识别模型

Plant leaf recognition model based on two-way convolutional neural network

【目的】针对卷积神经网络识别植物叶片过程中,叶片边缘形状对卷积层的过度作用而导致相似边缘形状叶片识别错误的问题,提出了一种双路卷积神经网络的植物叶片识别模型。【方法】模型考虑了叶片信息的边缘形状与内部纹理特征,构建了双路卷积神经网路结构,其中形状特征路径运用7层卷积层的网络结构,前3层采用大尺寸11×11及5×5的卷积核提取大视野特征,完成叶片形状特征提取,另外4层卷积层采用3×3小尺寸卷积核提取叶片细节特征;纹理特征路径采用6个3×3卷积核的卷积层,提取叶片纹理图像细节特征;然后通过特征融合层将两类特征相加为融合特征,并利用全连接层对植物叶片种类进行识别。【结果】实验结果表明,双路卷积神经网络模型与单路卷积神经网络和图像处理分类识别模型相比,在Flavia叶片数据集与扩充植物叶片数据集上,Top-1识别准确率分别提高到了99. 28%、97. 31%,Top-3识别准确率分别提高到了99. 97%、99. 74%,标准差较其他识别与分类模型下降到0. 18、0. 20。【结论】本文提出的叶片识别模型能有效避免相似叶片边缘形状干扰而导致识别错误的问题,可以提高植物叶片的识别准确率。

[Objective] Aiming at the problem that the leaf edge shape has an excessive effect on the convolution layer during the process of identifying the leaf of convolutional neural network, which leads to the error recognition of similar edge shape leaves, a plant leaf recognition model of two-way convolutional neural network was proposed. [Method] The model considers the edge shape and internal texture features of the blade information to construct a two-way convolutional neural network structure. Wherein, the shape feature path used a network structure of 7 layers of convolution layers, the first three layers used large-size 11×11 and 5×5 convolution kernels, extracting large field of view features to complete blade shape feature extraction, the other 4 layers of convolution layer used a 3×3 small size convolution core, extracting blade detail features. The two types of feature linear transformations were merged into one-dimensional feature vectors through a fully connected layer. Finally, the fully connected layer identified the plant leaf species. [Result] The experimental results showed that the two-way convolutional network model was compared with the single-channel convolutional network and the image recognition classification recognition model. On the Flavia leaf dataset and the expanded complex background leaf dataset, the accuracy of Top-1 recognition increased to 99.28% and 97.31%, respectively. The accuracy of Top-3 recognition increased to 99.97% and 99.74%, respectively. The standard deviation decreased to 0.18 and 0.20 compared with other identification and classification models. [Conclusion] The blade recognition and classification model proposed in this paper can effectively avoid the problems caused by the similar blade edge shape interference and improve the recognition accuracy of leaf plant species.