基于改进Vision Transformer网络的农作物病害识别方法  

Crop Disease Recognition Method Based on Improved Vision Transformer Network

作　　者：王杨 李迎春 许佳炜 王傲 马唱 宋世佳 谢帆 赵传信 胡明[2] WANG Yang;LI Yingchun;XU Jiawei;WANG Ao;MA Chang;SONG Shijia;XIE Fan;ZHAO Chuanxin;HU Ming(School of Computer and Information,Anhui Normal University,Wuhu 241002,China;School of Information and Artificial Intelligence,Wuhu Institute of Technology,Wuhu 241002,China)

机构地区：[1]安徽师范大学计算机与信息学院,安徽芜湖241002 [2]芜湖职业技术学院信息与人工智能学院,安徽芜湖241002

出　　处：《小型微型计算机系统》2024年第4期887-893,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61871412)资助;安徽省自然科学基金重点项目(KJ2019A0938,KJ2021A1314,KJ2019A0979)资助;安徽省社科规划基金项目(AHSKY2017D42)资助;安徽高校自然科学重点项目研究项目(KJ2017A552,KJ2019A0979,KJ2019A0511)资助;芜湖市科技计划项目(2021cg17)资助.

摘　　要：基于DCNN模型的农作物病害识别方法在实验室环境下识别准确率高,但面对噪声时缺少鲁棒性.为了兼顾农作物病害识别的精度和鲁棒性,本文在标准ViT模型基础上加入增强分块序列化和掩码多头注意力,解决标准ViT模型缺乏局部归纳偏置和视觉特征序列的自注意力过于关注自身的问题.实验结果表明,本文的EPEMMSA-ViT模型对比标准ViT模型可以更高效的从零学习;当添加预训练权重训练网络时,EPEMMSA-ViT模型在数据增强的PlantVillage番茄子集上能够得到99.63%的分类准确率;在添加椒盐噪声的测试数据集上,对比ResNet50、DenseNet121、MobileNet和ConvNeXt的分类准确率分别提升了6.08%、9.78%、29.78%和12.41%;在添加均值模糊的测试数据集上,对比ResNet50、DenseNet121、MobileNet和ConvNeXt的分类准确率分别提升了18.92%、31.11%、20.37%和19.58%.The crop disease recognition method based on Deep Convolutional Neural Networks Models have high identification accuracy in laboratory environment,but lacks robustness in the face of noise.In order to take into account the recognition accuracy and robustness of crop disease recognition,this paper adds Enhanced Patch Embedding and Masked Multi-head Self-Attention to the Vision Transformer model,solving the problem of lacking local inductive bias and the self-attention of visual feature sequences being too focused on itself of ViT model.The experimental results show that the EPEMMSA-ViT model in this paper can learn from scratch more efficiently than ViT model;when adding pre-training weights to train the network,the classification accuracy obtains 99.63%on the augmented PlantVillage tomato subset;when adding salt and pepper noise on the test data set,the classification accuracy of the EPEMMSA-ViT model compared with ResNet50,DenseNet121,MobileNet and ConvNeXt increased by 6.08%,9.78%,29.78%and 12.41%respectively;on the test data set with blur noise,the EPEMMSA-ViT model compared with ResNet50,DenseNet121,MobileNet and ConvNeXt,the classification accuracy of is improved by 18.92%,31.11%,20.37%and 19.58%respectively.

关 键 词：农作物病害识别 深度卷积神经网络 视觉Transformer 自注意力 局部归纳偏置 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]