结合小目标感知与边缘部署的烟叶病害检测方法

Tobacco Disease Detection Method Integrating Small-target Perception and Edge Deployment

摘要:

针对现有检测方法在烟叶病害图像上表现出小目标病斑易漏检、特征提取能力不足及难以在边缘设备部署的问题，本文提出了一种基于改进YOLO 11s的烟叶病害检测算法PCG-YOLO。使用无人机低空拍摄构建了5种烟叶病害图像数据集;为解决小面积病斑检测难的问题，引入P2小目标检测层（P2 small-object detection layer），显著增强了模型对浅层特征的利用能力;针对YOLO 11s在特征融合阶段对关键区域关注不足的问题，设计了C3k2_RVS模块（C3k2_RepViTBlock_SimAM），其中RepViTBlock提升了计算效率，SimAM 3D注意力机制帮助模型更精准聚焦于病斑区域，在提升检测精度的同时控制了计算成本;最后，针对模型深层传统卷积模块参数量较高，采用轻量化GSConv模块缓解高维度特征图产生的计算负担，有效降低了模型复杂度。试验结果表明，PCG-YOLO精确率、mAP50和mAP50-95分别达到83.3%、92.4%和77.1%，相较于基准模型分别提升4.2、2.7、3.6个百分点，检测速度达到64.1 f/s;多模型对比试验结果表明，PCG-YOLO模型在检测精度和检测速度之间取得了较好的平衡，整体性能优于主流目标检测算法。将PCG-YOLO部署到Jetson TX2边缘计算设备上，经TensorRT加速优化后精确率、mAP50和mAP50-95分别为82.9%、90.6%和76.3%，检测速度达到27.4 f/s。本文方法对烟叶病害图像具有一定识别能力，并且能够在低功耗边缘计算设备上稳定运行。

Abstract:

Aiming to address the challenges of small-target omission, insufficient feature extraction, and edge deployment difficulty in tobacco leaf disease detection, PCG-YOLO was proposed based on an improved YOLO 11s framework. A five-class tobacco disease dataset was constructed by using low-altitude UAV imaging. A P2 small-object detection layer was added to enhance shallow feature utilization. A customized C3k2_RVS module (C3k2_RepViTBlock_SimAM) was designed, where Rep_ViTBlock improves computational efficiency and SimAM enabled precise attention to lesion areas. Additionally, the lightweight GSConv module reduced model complexity. Ablation results showed that the P2 layer increased mAP50-95 by 4.7 percentage points, accompanied by a speed reduction of 14.8 f/s. GSConv reduced parameters by 13.8%, while C3k2_RVS enhanced both accuracy and speed. The proposed PCG-YOLO achieved a precision of 83.3%, an mAP50 of 92.4%, and an mAP50-95 of 77.1%, improving over the baseline by 4.2, 2.7, and 3.6 percentage points, respectively, with a detection speed of 64.1 f/s. Multi-model comparisons indicated that the proposed method achieved the best balance between accuracy and speed, outperforming others in mAP metrics. Detr-ResNet18 yielded higher precision at 84.2%, but suffered from slower speed and higher computational cost. YOLO v5s was faster and more lightweight, yet less accurate. Finally, the proposed method was deployed on a Jetson TX2 with TensorRT acceleration, achieving a precision of 82.9%, an mAP50 of 90.6%, an mAP50-95 of 76.3%, and a speed of 27.4 f/s, demonstrating strong recognition capability and stable operation on low-power edge platforms.