【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。更多还原

【Abstract】 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.更多还原