面向边缘计算的水稻病害检测方法与装置研究

Development of Rice Disease Detection Methods and Devices for Edge Computing

摘要:

针对自然环境下水稻病害检测准确性与效率不高等问题，本文提出一种改进YOLO v5s的轻量化检测模型YOLO-RD，并将其部署至边缘计算设备，设计便携式检测装置，实现水稻病害快速检测。通过引入GhostNet网络减少计算量和参数量。结合轻量级注意力机制Shuffle Attention和动态检测头DyHead(Dynamic head)，增强模型对水稻病害图像的特征提取和自适应检测能力。使用Shape-IoU替代CIoU损失函数，提升自然环境下检测精度。试验结果表明，YOLO-RD模型在平均精度均值(mean Average Precision, mAP)达到94.2%的同时，大幅降低计算复杂度和参数量，具有良好的轻量化效果，与基准模型相比计算量、参数量和内存占用量分别减少44.4%、43.2%和41.3%。与YOLO 11n、YOLO v8n和YOLO v5n等目标检测模型对比，本模型检测效果最佳。将模型在树莓派4B边缘计算设备上部署，单幅图像检测时间为1.97s，满足实际应用需求，可为水稻病害智能化检测提供高效可行的解决方案。

Abstract:

Aiming to address the challenges of accuracy and efficiency in rice disease detection under natural conditions, a lightweight detection model, YOLO-RD, was presented based on an improved YOLO v5s framework. The model was optimized and successfully deployed on edge computing devices, enabling the creation of a portable device for fast rice disease detection. In the proposed model, GhostNet was integrated to reduce computational complexity and the number of parameters. Finally, the lightweight Shuffle Attention mechanism and the dynamic detection head DyHead were employed to enhance feature extraction and adaptive detection capabilities, particularly for complex disease features. Furthermore, the standard CIoU loss function was replaced by Shape-IoU to improve detection performance in challenging environments by focusing on shapebased regression. Experimental results demonstrated that YOLO-RD achieved a mean Average Precision (mAP) of 94.2%, while significantly reducing computational complexity and parameter size. Specifically, YOLO-RD reduced computation, parameters, and weight by 44.4%, 43.2%, and 41.3%, respectively, compared with the baseline model. In addition, the model outperformed detection models such as YOLO 11n, YOLO v8n, YOLO v5n, and others in terms of accuracy. When deployed on a raspberry Pi 4B edge computing device, YOLO-RD achieved an inference time of 1.97 s per image, meeting the requirements for real-time application. These findings suggested that YOLO-RD offered an efficient and robust solution for intelligent rice disease detection in practical agricultural scenarios.