基于改进的ConvNeXt网络的茶叶病虫害分类识别方法研究

Research on Tea Pest and Disease Classification Based on an Improved ConvNeXt Network

摘要: 随着茶叶种植规模的持续扩大，病虫害对产量与品质的影响日益显著，已成为制约茶产业发展的重要因素之一。由于不同病虫害类型对应的防治措施存在差异，实现对其类别的准确判别对于开展针对性防控具有关键意义。传统方法多依赖人工经验或浅层特征，存在识别效率低、准确性不足等问题。近年来，卷积神经网络(CNN)凭借其优异的特征学习能力，在病虫害图像识别领域得到广泛应用。因此，本文提出了一种改进的ConvNeXt茶叶病虫害分类模型。在原有ConvNeXt Block中引入融合空间与通道信息的Dual注意力机制，形成DW-Dual结构，以增强特征表达能力；在此基础上进一步嵌入SCConv模块，构建SCConvNeXt-Dual结构，从而强化特征建模能力并提升训练效率。实验结果表明，改进模型在收敛速度与分类性能方面均优于基准模型，收敛所需迭代次数由约140轮降低至约70轮，分类准确率由93.7%提升至95.4%。综上，本文实现了提升茶叶病虫害识别精度的同时兼顾训练效率，为相关领域的智能化识别与精准防控提供了有益参考。

Abstract: With the continuous expansion of tea cultivation, pest and disease issues have become increasingly prominent, posing significant threats to both yield and quality and thus constraining the sustainable development of the tea industry. Due to the variation in control strategies for different types of pests and diseases, accurate identification of their categories is essential for implementing targeted management measures. Traditional approaches, which rely heavily on manual experience or shallow features, often suffer from low efficiency and limited accuracy. In recent years, convolutional neural networks (CNNs) have been widely applied in pest and disease recognition tasks owing to their powerful feature extraction capabilities. In this study, an improved ConvNeXt-based model for tea pest and disease classification is proposed. Specifically, a Dual attention mechanism integrating spatial and channel attention is introduced into the ConvNeXt Block to construct the DW-Dual module, thereby enhancing feature representation. Furthermore, the SCConv module is embedded into the DW-Dual structure to form the SCConvNeXt-Dual module, which strengthens feature modeling capability and improves training efficiency. Experimental results demonstrate that the proposed model outperforms the baseline ConvNeXt in both convergence speed and classification performance. The number of training epochs required for convergence is reduced from approximately 140 to 70, while the classification accuracy is improved from 93.7% to 95.4%. In summary, the proposed method achieves higher recognition accuracy while maintaining efficient training, providing a useful reference for intelligent pest and disease identification and precise control in tea cultivation.