基于轻量级密集多尺度注意力网络的小麦叶部锈病识别方法

Lightweight Dense Multi-scale Attention Network for Identification of Rust on Wheat Leaves

摘要:

人工诊断小麦锈病成本高、效率低，已无法满足现代农业生产的需要。本文提出了一种轻量级密集多尺度注意力网络模型(Mobile-Dense multi-scale attention net, Mobile-DMSANet)，用于自动识别田间自然场景中的小麦叶部锈病（条锈病和叶锈病）。该模型在输入层设计了一个快速下采样模块(Fast subsampling block, FSB)，它在不增加计算成本的前提下提高模型的特征表达能力。模型的特征提取层使用3个轻量级特征提取模块(Dense multi-scale attention, DMSA)来提取小麦叶部锈病的特征。DMSA模块设计了一个多尺度的3路卷积层(Multi-scale three-way convolution, MSTC)用于获得不同尺度感受野，以提高模型的表达能力和对不同尺寸锈病的感知能力。DMSA模块中6个MSTC层通过密集连接实现特征重用，不仅大大减少了模型的参数量，而且提高了对这两种相似的小麦叶部锈病的特征提取能力。在DMSA模块中还引入了协调注意力机制(Coordinated attention, CA)，来提高对病害信息的敏感性，并抑制图像中的背景信息。模型的输出层使用Softmax函数实现小麦叶部锈病识别。结果表明，Mobile-DMSANet模型在测试数据集上的识别准确率为96.4%，高于经典CNN模型(如ResNet50、AlexNet)和轻量级CNN模型(如ShufflenetV2、DenseNet系列)。Mobile-DMSANet参数量为4.54×105，与其他轻量级模型相比大幅下降。本文所设计模型可用于移动端小麦叶部锈病的自动识别。

Abstract:

Artificial identification of wheat rust is costly and inefficient, and can no longer meet the needs of modern agricultural production. A lightweight dense multi-scale attention network model called Mobile-DMSANet was presented for the automatic identification of rust on wheat leaves (stripe rust and leaf rust) from images of natural scenes taken in the field. In the input layer of the network, a fast subsampling block (FSB) was used to improve the feature expression ability of the network without adding computational cost. In the feature extraction layer, three lightweight blocks called dense multi-scale attention (DMSA) blocks were used to extract the features of rust on wheat leaves. In the DMSA block, a multi-scale three-way convolution (MSTC) layer was designed to get different scales for the receptive fields, in order to improve the expressive ability of the network and its ability to perceive the features of rust disease at different scales. Six MSTC layers were used to achieve feature reuse by dense connections in the DMSA block, an approach that not only greatly reduced the number of parameters of the network but also improved the feature extraction ability for similar diseases. A coordinated attention (CA) block was also introduced to the DMSA block to increase the sensitivity to positional information and suppress background information in the image. The output layer of the network used a Softmax function to classify rust on wheat leaves. The results showed that the recognition accuracy of Mobile-DMSANet model on the test dataset was 96.4%, which was higher than that of other models. Mobile-DMSANet had only 454000 parameters, less than for other lightweight models. The proposed model can be used for the automatic identification of rust on wheat leaves using mobile devices.