基于可见光谱和改进注意力的农作物病害识别  

Crop Disease Recognition Based on Visible Spectrum and Improved Attention Module

作　　者：孙文斌 王荣 高荣华[1,3] 李奇峰[1,3] 吴华瑞[1,3] 冯璐 SUN Wen-bin;WANG Rong;GAO Rong-hua;LI Qi-feng;WU Hua-rui;FENG Lu(Beijing Research Center for Information Technology in Agriculture,Beijing 100097,China;College of Information Engineering,Guangdong University of Technology,Guangzhou 510006,China;National Engineering Research Center for Information Technology in Agriculture,Beijing 100097,China;College of Information Engineering,Northwest Agriculture and Forestry Univer sity,Yangling 712100,China)

机构地区：[1]北京农业信息技术研究中心,北京100097 [2]广东工业大学信息工程学院,广东广州510006 [3]国家农业信息化工程技术研究中心,北京100097 [4]西北农林科技大学信息工程学院,陕西杨凌712100

出　　处：《光谱学与光谱分析》2022年第5期1572-1580,共9页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金面上项目(61771058);北京市科技计划课题(Z191100004019007)资助。

摘　　要：基于可见光谱的农作物病害自动化识别和诊断是一个具有挑战性的研究领域,但现有基于卷积神经网络进行病害识别的研究往往利用深层网络牺牲模型参数量来提高对单一农作物病害识别的准确率,从而造成硬件资源的浪费。为提高农作物病害识别的准确率且避免深层网络的使用,该研究将注意力机制引入农作物病害识别领域,提出了一种基于可见光谱和改进注意力机制的浅层农作物病害识别模型,设计了新的注意力模块SMLP和农作物病害识别模型SMLP_ResNet。利用卷积层代替全连接层设计参数量较少的残差网络(ResNet),然后利用SMLP、归一化结构(BatchNormalization)和残差模块(Res_block)得到改进的残差模块单元(SMLP_Res),其中SMLP由全局池化和多层感知机组成,建立各通道间依赖关系。多层感知机使用三层网络结构,将全局特征的通道维度提升至两倍,然后对其通道维度进行两次降维,恢复至原始维度,减少了全局特征损失。SMLP_Res可对通道中的病害特征重校准,减少对识别任务无效的冗余信息,最后构建农作物病害识别模型SMLP_ResNet,在减少模型层数同时提高其识别率。使用两个不同难度的多种植物和病害混合的公开数据集AIChallenger2018和PlantVillage验证本文模型。实验结果表明,SMLP_ResNet模型在18、50和101层时达到了较高的识别率,其中SMLP_ResNet18模型效果最佳,在两个数据集中的病害识别率分别为86.93%和99.32%。SMLP_ResNet18的准确率不仅高于改进前的ResNet18和SENet18网络,还高于其他研究者提出的模型的准确率,且模型权重大小为48.6MB,仅约为AlexNet网络权重的五分之一,能够在模型参数量较小的情况下实现较高的病害识别率。从Grad-CAM生成的热力图中可看出SMLP_ResNet18相比于其他模型更关注于叶片病害部位的特征,其背景信息以及叶片健康部位的权值较小。该研究所提出的SMLP_ResNeAutomatic identification and diagnosis of crop diseases based on visible spectrum is a challenging research field that agricultural sectors have given great attention.However,the existing research of disease identification based on convolutional neural network is prone to sacrifice network depth for the results of single disease detection,which usually resulting in the waste of hardware resources.This paper proposes a low-level crop disease identification model based on visible spectrum and improved attention mechanism,which designs a novel attention module SMLP(Squeeze-Multi-Layer-Perceptron)and crop disease identification model SMLP_ResNet.This network first uses a convolutional layer to replace the full connect layer to construct a less-parameter residual network(ResNet)whilst improving residual block(SMLP_Res)by the SMLP,Batch Normalization(BN)module and residual block(Res_block).The SMLP consists of global pooling layers and multi-layer perceptron,which can help to build a dependent relationship between channels.The multi-layer perceptron utilizes a three-layer network to double the channel dimension of global features,then restores the channel to the original dimension through carrying out a twice dimensionality reduction operation to minimize the loss of global features.On the above basis,the SMLP_Res can recalibrate the features of disease and reduce the redundant information that is useless for the detection target,and finally constructs a crop disease identification model SMLP_ResNet reduces the number of layers and enhances the accuracy of identification at the same time.Moreover,the proposed network is verified on two datasets with multiple crops and diseases,AI Challenger 2018 and Plant Village.The experimental results show that the SMLP_ResNet model achieves a high classification accuracy at the 18,50 and 101-layer of the network,especially SMLP_ResNet18 has the best performance with accuracies up to 86.93%and 99.32%on two verification datasets respectively.SMLP_ResNet18 not only performs better tha

关 键 词：病害识别 图像分类 注意力机制 残差网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]