针对传统水稻病虫害识别方法存在的效率低下、易受主观因素干扰等局限性，以及现有深度学习模型在捕捉水稻病虫害细微特征和处理类别不平衡数据方面存在的不足，开展了一系列研究与实践。首先，利用AR眼镜在水稻田间实地采集病虫害图像，结合公开数据集IP102和网络上的图像构建了水稻病虫害数据集，并采用数据增强方式扩充训练样本，缓解类别不平衡与图像质量问题；其次，在EfficientNet-V2模型的基础上，引入CBAM注意力机制替换原有的SE模块以增强对水稻病虫害的细节特征捕捉，并采用PolyLoss损失函数优化不平衡数据学习，构建了EfficientNet-V2-Rice水稻病虫害识别模型；最后，基于改进后的识别模型，开发了一款配套的安卓手机端智能识别APP。该APP功能丰富，集成了用户注册与登录、图像上传、智能识别、识别结果检索与详情查看等核心功能模块。用户只需通过手机摄像头拍摄水稻病虫害图像并上传至APP，即可快速获得准确的识别结果，并可随时检索和查看历史识别记录的详细信息。为验证模型改进策略的有效性，进行了消融实验和对比实验。实验结果表明，本文提出的EfficientNet-V2-Rice模型在水稻病虫害识别任务中表现优秀，精确率、召回率和F₁分数分别达到84.92%、86.00%和85.45%。基于此模型开发的安卓手机端APP，为用户提供了便捷高效的识别服务，为水稻病虫害的智能监测与辅助诊断提供了一种实用的工具。

To address the limitations of traditional rice pest and disease recognition methods, such as low efficiency and susceptibility to subjective interference, as well as the shortcomings of existing deep learning models in capturing subtle features of rice pests and diseases and handling category-imbalanced data, a series of research and practical efforts have been undertaken. Firstly, pest and disease images were collected in rice fields using AR glasses. A rice pest and disease dataset was then constructed by combining these images with the publicly available dataset IP102 and images sourced from the web. Data augmentation techniques were employed to expand the training samples, thereby mitigating issues related to category imbalance and image quality. Secondly, based on the EfficientNet-V2 model, the CBAM (Convolutional Block Attention Module) attention mechanism was introduced to replace the original SE (Squeeze-and-Excitation) module, aiming to enhance the model’s ability to capture detailed features of rice pests and diseases. Additionally, the PolyLoss loss function was adopted to optimize the learning process for unbalanced data, leading to the construction of the EfficientNet-V2-Rice model for rice pest and disease recognition. Finally, leveraging the improved recognition model, a companion intelligent recognition APP for Android smartphones was developed. This APP boasts a rich set of features, integrating core modules such as user registration and login, image uploading, intelligent recognition, retrieval of recognition results, and viewing of detailed information. Users can simply capture images of rice pests and diseases using their smartphone cameras and upload them to the APP to quickly obtain accurate recognition results. They can also retrieve and view detailed information about historical recognition records at any time. To verify the effectiveness of the model improvement strategy, ablation and comparison experiments were conducted. The experimental results demonstrate that the proposed EfficientNet-V2-Rice model performs exceptionally well in rice pest and disease recognition tasks, achieving precision, recall, and F₁ scores of 84.92%, 86.00%, and 85.45%, respectively. The Android smartphone APP developed based on this model provides users with convenient and efficient recognition services, offering a practical tool for the intelligent monitoring and auxiliary diagnosis of rice pests and diseases.