基于注意力机制与EfficientNet的轻量化水稻病害识别方法

为实现水稻病害图像的快速、准确识别，提出一种基于注意力机制与EfficientNet的轻量化水稻病害识别方法。该方法首先引入轻量级卷积注意力模块(Convolutional Block Attention Module，CBAM)改进Efficientnet-B0中的主体模块轻量翻转瓶颈卷积核(Mobile Inverted Bottleneck Convolution，MBConv)，然后利用Ghost模块优化网络中的卷积层，降低网络的参数量和计算量，最后使用Adam优化算法提高网络的收敛速度。在由572幅水稻白叶枯病、稻粒黑粉病、稻曲病、稻胡麻斑病和健康叶片5类水稻图像构成的测试集上，本文所提方法的识别准确率为95.63%，较EfficientNet-B0提高1.75%；分别比同类经典神经网络VGG16、Inception-V3、ResNet101和DenseNet201提高8.39%、4.72%、3.67%和1.05%。本文所提方法模型参数量为4.4 M，较EfficientNet-B0减少2.8 M；相比于对照网络，其参数量仅是这些网络模型参数量的9.05%、18.37%、9.81%和21.64%。试验结果表明：本文所提方法能够实现对不同水稻病害图像的准确、快速识别，而且识别模型轻量，具有较少的网络参数量。

Lightweight rice disease identification method based on attention mechanism and EfficientNet

To achieve fast and accurate recognition of rice disease images, a lightweight rice disease recognition method based on an attention mechanism and EfficientNet is proposed. We used the convolutional block attention module as the lightweight attention module to improve the mobile inverted bottleneck convolution of the main module in EfficientNet-B0. Then the Ghost module is used to optimize the convolution layer of the network to reduce the number of parameters and computation time of the network. Finally, the Adam optimization algorithm is used to improve the convergence speed of the network. On the test set, which contained 572 rice images of rice bacterial blight, rice kernel smut, rice false smut, rice helminthosporium leaf spot, and healthy leaves, the recognition accuracy of the proposed method was 95.63%. This was 1.75% higher than the recognition accuracy of EfficientNet-B0 on the same test set, and 8.39%, 4.72%, 3.67%, and 1.05% higher than the recognition accuracy of similar classical neural networks VGG16, Inception-V3, ResNet101, and DenseNet201, respectively. The number of parameters of the proposed lightweight rice disease recognition method was 4.4 M, which was 2.8 M less than that of EfficientNet-B0. Compared with the comparison network, the number of parameters of the proposed method was 9.05%, 18.37%, 9.81%, and 21.64% of them. Our results show that the proposed method can identify different rice disease images accurately and quickly.