Abstract:  With the close combination of deep learning technology and agriculture, more researches have applied deep learning technology to recognize agricultural diseases. Agriculture with deep learning improves the yield and quality of agricultural products. This paper proposes a new Xceptionbased method for the identification of plant diseases. Considering the image of plant diseases will be disturbed by uncertain environmental factors and reduce the image information on the basis of Xception, a new Channel Expand Module (CEM) module adopting the combination of Multiscale Depthwise Separable Convolution with channel distribution weight and group convolution was proposed. This enhanced the feature extraction efficiency of space and channel as well as the new strategy of channel expansion which included channel expansionmaintainreexpansioncompression to further optimize the channel feature extraction in the network. The novel dense connection was proposed to improve feature reuse among feature maps of the same size. The experimental data set consisted of 50 types of images of 10 different plants, including 10 healthy plants and 27 diseases,  which 13 diseases were classified into two levels. The algorithm obtained 91.9% accuracy, 88.7% precision, 8245% recall, and 85.33% F1Score. Compared with Xception architecture, our network had a smaller model complexity with FLOPs being 29.33 M, which was 66.4% of Xception while smaller parameter amount with parameter amount being 14.05 M, which was 66.9% of Xception. Therefore, XceptionCEMs can effectively identify diseases, which is conducive to the development of intelligent agriculture.

Key words: deep learning, plant disease, feature extraction, dense connection, image recognition

摘要： 随着深度学习技术与农业的密切融合,越来越多的研究将深度学习技术用于农业病虫害检测,提高农产品产量和质量。本文提出一种新颖的基于Xception模型的植物病害识别方法。了解到植物病害图像会受到不确定环境因素的干扰而减小图像信息。在Xception的基础上,提出一种新的通道扩增模块,采用带有通道分配权重的多尺度深度卷积与组卷积结合,增强空间和通道的特征提取效率;在网络中采用通道扩张-保持-再扩张-压缩的新策略,进一步优化通道特征提取;引入密集连接方式,提高在同尺寸的特征图之间特征重用。试验数据集由10种不同植物的50类图像组成,分别包括10种健康植物和27种病害,其中对13种病害进行了两种程度的分类。本文的方法在这些类别上可以获得91.9%的准确率,88.7%的精确率,82.45%的召回率以及85.33%的F1值。本文的算法有更小的模型复杂度和参数量,计算量为29.33 M,为Xception的66.4%。参数量为14.05 M,为Xception的66.9%。因此,Xception-CEMs能够有效对病虫害进行识别,有利于农业智能化发展。

关键词: 深度学习, 植物病害, 特征提取, 密集连接, 图像识别