浅谈CNN在柑橘病虫害识别预警中的应用

A Brief Discussion on the Application of CNN in Citrus Disease and Pest Identification and Early Warning

摘要: 基于深度可分离卷积与CBAM注意力机制协同优化的改进ResNet34模型，在柑橘病虫害智能识别任务中实现了高精度、高效率与轻量化的统一。该模型通过引入深度可分离卷积有效降低了计算复杂度与参数量，同时结合CBAM模块的通道与空间注意力机制，自适应聚焦病虫害关键特征区域，显著提升了模型的表征能力与判别力。实验表明，该识别系统在自建柑橘病虫害数据集上平均精确度可达96%，在保证实时性的同时，大幅压缩了模型体积，为移动端与边缘计算设备部署提供了可行的技术方案，具备较强的实际应用价值与推广潜力。

Abstract: Based on the collaborative optimization of depthwise separable convolution and CBAM attention mechanism, the improved ResNet34 model achieves a unification of high precision, high efficiency, and lightweight in the intelligent recognition task of citrus diseases and pests. By introducing depthwise separable convolution, the model effectively reduces computational complexity and the number of parameters, while the channel and spatial attention mechanisms of the CBAM module adaptively focus on key feature regions of diseases and pests, significantly enhancing the model’s representational capability and discriminative power. Experiments demonstrate that the recognition system achieves an average precision of up to 96% on a self-built citrus disease and pest dataset. While ensuring real-time performance, the model size is substantially compressed, providing a feasible technical solution for deployment on mobile and edge computing devices, with strong practical application value and promotion potential.