基于Transformer和CNN的真实场景下植物病害识别方法  

Plant disease identification method in real scenarios based on Transformer and CNN

作　　者：刘畅 莫海芳[1,3] 马春 Liu Chang;Mo Haifang;Ma Chun(School of Computer Science,South-Central MinZu University,Wuhan 430074,China;Hubei Provincial Engineering Research Center of Agricultural Blockchain and Intelligent Management,Wuhan 430074,China;Hubei Provincial Engineering Research Center for Intelligent Management of Manufacturing Enterprise,Wuhan 430074,China)

机构地区：[1]中南民族大学计算机科学学院,武汉430074 [2]农业区块链与智能管理湖北省工程研究中心,武汉430074 [3]湖北省制造企业智能管理工程技术研究中心,武汉430074

出　　处：《现代计算机》2023年第11期22-27,共6页Modern Computer

基　　金：国家民委中青年英才培养计划(MZR20007);新疆维吾尔自治区区域协同创新专项(科技援疆计划)(2022E02035)。

摘　　要：针对真实场景下的植物病害图像,设计了一种融合Transformer和CNN的植物病害识别模型CLT。该模型采用多阶段层次设计,以Conv Stem作为初始特征提取模块提取浅层局部特征,结合Transformer模块以学习全局特征;利用卷积Token嵌入,改变模型每个阶段Token的序列长度和特征维数,实现对多层次的局部空间上下文进行建模;将卷积模块融入到Transformer模块,并将多头注意力机制中的线性投影替换为卷积投影,提升模型对局部空间的特征提取。实验结果表明,CLT模型能够有效表达植物病害的各种特征,在真实场景下平均准确率达到77.91%,分类效果优于其他模型,为真实场景下植物病害识别提供参考。A plant disease recognition model CLT incorporating Transformer and CNN is designed for plant disease images in real scenarios.the model adopts a multi-stage hierarchical design,using Conv Stem as the initial feature extraction module to extract shallow local features,combined with Transformer module to learn global features;using convolutional Token embedding to change the model The Convolutional Token embedding is used to change the sequence length and feature dimension of Token at each stage to model the multi-level local space context;the Convolutional module is incorporated into the Transformer module and the linear projection in the multi-headed attention mechanism is replaced with the Convolutional projection to enhance the model’s feature extraction of the local space.The experimental results show that the CLT model can effectively express various features of plant diseases,with an average accuracy of 77.91%in real scenes,and the classification effect is better than other models,providing a reference for plant disease recognition in real scenes.

关 键 词：植物病害识别 TRANSFORMER CNN 真实场景 

分 类 号：S432[农业科学—植物病理学] TP183[农业科学—农业昆虫与害虫防治] TP391.41[农业科学—植物保护]