Abstract: To address the problems of low efficiency, high cost, and low accuracy in manual detection of citrus diseases, this article combines artificial intelligence technology to classify and identify diseased citrus leaves. Firstly, a dataset of citrus disease leaves under simulated complex environments is established. Secondly, an improved Swin Transformer model for citrus disease leaf classification is proposed, which includes a Local Perception Channel Enhanced Attention Module (LPCE) to enhance the models receptive field and feature representation capabilities. Through weighted correlation between channels, the model is made to extract key features more easily. Experiments demonstrate that the classification accuracy of the proposed model reaches 98.52%, with Precision, Recall, and F1score reaching 98.17%, 98.24%, and 98.28% respectively, all exceeding the baseline model. It providing technical support for the detection of citrus diseases.

Key words: citrus diseased leaves, deep learning, classification recognition, SwinTransformer, attention module

摘要： 针对柑橘病害人工检测效率低、成本高、准确度低等问题，结合人工智能技术对柑橘病叶进行分类识别。首先，建立模拟复杂环境下的柑橘病叶数据集。其次，提出一种改进的SwinTrasnformer柑橘病叶分类模型，包含局部感知通道增强注意力模块(LPCE)，以提升模型的感受野和特征表达能力，通过通道之间的相关性进行加权，使模型更容易提取关键特征。试验证明本文模型的分类识别准确率达到98.52%，精确率、召回率和F1score分别达到98.17%、98.24%、98.28%，均超过基线模型。该模型为柑橘病害的检测提供技术支撑。

关键词: 柑橘病叶, 深度学习, 分类识别, Swin-Transformer, 注意力模块