基于残差注意力网络的马铃薯叶部病害识别

Potato leaf disease recognition via residual attention network

马铃薯叶部病害严重制约着马铃薯的产量,为此提出了一种基于注意力和残差思想的深度卷积神经网络模型RANet。依据注意力机制,在RANet中构建并行池化的注意力模块,以增强网络的特征提取能力,并借助残差思想避免注意力模块造成的特征值衰减。以早疫病初期、早疫病晚期、晚疫病初期、晚疫病晚期和健康叶片的叶部图像为研究对象,RANet的平均识别率为93.86%,比ResNet50、VGG16、ShuffleNet和MobileNet高2.46%~16.13%。通过对注意力模块参数量的控制,使该模型图像识别速度可达73 ms/张。

Considering the serious impact of potato leaf diseases on the production of potatoes,a deep convolutional neural network model RANet based on attention and residual thought was proposed.Based on attention mechanism,a parallel pooled attention module was constructed in RANet to enhance the feature extraction ability of the network,and the feature value decay caused by the attention module was restricted by using the residual thought.Taking the leaves in the early period of early blight,late period of early blight,early period of late blight,and late period of late blight and healthy leaf images as the research object,the average recognition rate of RANet was found to be 93.86%,which was 2.46%~16.13%higher than ResNet50,VGG16,ShuffleNet and MobileNet.By controlling the parameters of the attention module,the recognition speed of the model could be achieved to reach 73 ms per image.