基于深度学习的柑橘黄龙病症状识别

Citrus Huanglongbing symptom recognition based on deep learning

【目的】黄龙病是全球柑橘产业面临的重大挑战,由于其潜伏期长和初期症状不明显,早期发现与防治成为关键。传统的人工诊断方法耗时且效率低下,容易错过最佳防治时间。随着技术进步,卷积神经网络(CNN)在农作物病虫害检测中展现出巨大潜力,将CNN引入Transformer也正在引起研究者的兴趣。基于卷积视觉交换器(convolutional vision transformer,CvT)的改进模型识别柑橘黄龙病症状并分类。【方法】研究采集并构建了柑橘叶片数据集,利用数据增强技术扩充训练样本,并设计了轻量级多头自注意力模块(LMHSA)和反向残差前馈神经网络(IRFFN)模型,以提高模型的泛化能力和病症识别的准确性。【结果】改进后的CvT模型在柑橘黄龙病的检测任务中表现优异,分类精度为97.6%,可以实现对柑橘黄龙病的精确识别。【结论】该模型有可能成为一种柑橘黄龙病快速鉴定的辅助工具,后面将进一步优化该模型的精度和可靠性。

[Objective]Citrus Huanglongbing(HLB)poses a significant challenge to the global citrus industry,and so the early detection and control become crucial due to its long latency period and inconspicuous initial symptoms.Traditional manual diagnostic methods are time-consuming,inefficient,and prone to missing the optimal treatment window.With technological advancements,Convolutional Neural Networks(CNN)have shown great potential in crop disease and pest detection,and the integration of CNN into transformers has also garnered interest among researchers.[Method]In this study,an improved model based on the CvT(Convolutional vision Transformer)model was developed to identify and classify symptoms of citrus Huanglongbing.A citrus leaf dataset was collected and constructed,and data augmentation techniques were employed to expand the training samples.Lightweight Multi-Head Self-Attention(LMHSA)and Inverse Residual Feedforward Neural Network(IRFFN)models were designed to enhance the model's generalization ability and accuracy in symptom recognition.[Result]Experimental results demonstrated that the improved CvT model performed exceptionally well in the detection of citrus Huanglongbing,achieving a classification accuracy of 97.6%and enabling precise identification of the disease.[Conclusion]This model has the potential to become an auxiliary tool for the rapid identification of citrus Huanglongbing,and the accuracy and reliability of the model will be further optimized in the future.