基于轻量化YOLO的水稻病害检测研究

Research on Rice Disease Detection Based on Lightweight YOLO

摘要: 针对现有水稻病害检测模型参数量大、硬件部署难度高的问题，本文基于YOLO11提出一种轻量化改进算法YOLO-SGHM。该模型采用Ghost-HGNetv2骨干网络替代原有网络结构，通过Ghost模块生成冗余特征图以降低计算量；颈部网络融入C3k-Star模块，提升特征融合能力；在检测头卷积层中实施信息共享策略，使模型参数量降低25.7%；引入最小点距离交并比(MPDIoU)损失函数优化边界框回归效果。实验结果表明，改进模型相较于YOLO11，参数量降至1.24 M，降幅达52.1%，计算量降低33.3%至4.2 GFLOPs；在含6960帧样本的水稻病害数据集上，模型平均精度均值(mAP)达96.3%，帧率提升14.6%。该研究为移动植保设备提供了一种高效的病害检测方案。

Abstract: In view of the large parameter size and high difficulty in hardware deployment of existing rice disease detection models, this paper proposes a lightweight improved algorithm YOLO-SGHM based on YOLO11. The model adopts the Ghost-HGNetv2 backbone network to replace the original structure, and the Ghost module generates redundant feature maps to reduce the computational load. The neck network integrates the C3k-Star module to enhance the feature fusion ability. By implementing the information sharing strategy in the convolutional layer of the detection head, the model reduces the parameters by 25.7%. The Minimum Point Distance IoU (MPDIoU) loss function is introduced to optimize the bounding box regression effect. Experimental results show that compared with YOLO11, the improved model reduces the parameter size to 1.24 M with a decrease of 52.1%, and the computational load is reduced by 33.3% to 4.2 GFLOPs. On the rice disease dataset with 6960 frames, the model achieves a mean average precision (mAP) of 96.3% and the frame rate is increased by 14.6%. This research provides an efficient disease detection solution for mobile plant protection equipment.