基于改进YOLO 11与自适应通道DepGraph剪枝的轻量化水稻病虫害检测方法

Lightweight Rice Pest and Disease Detection Method Using Improved YOLO 11 with Adaptive Channel-wise FC-DepGraph Pruning

摘要:

针对当前水稻病虫害深度学习识别模型参数量大、计算复杂度高，难以在资源受限的农业嵌入式终端实现实时检测的工程难题，本文提出一种面向边缘端部署的轻量化水稻病虫害检测方法YOLO 11-DG。以YOLO 11n为基线模型，设计融合FlexConv通道动态匹配算法的FC-C3K2模块，有效解决了传统C3K2模块在结构化剪枝过程普遍存在的通道维度失配问题;在损失函数层面，提出引入归一化距离惩罚项的边界框回归损失函数YIoU，显著缓解了CIoU损失在预测框与真实框高度重合时出现的梯度消失缺陷。在此基础上，引入FC-DepGraph结构化剪枝方法构建模型轻量化框架，实现了剪枝过程跨模块通道的自动对齐，大幅提升了剪枝效率与模型压缩性能。试验结果表明，YOLO 11-DG mAP50达到84.5%，参数量仅2.36×10^6，浮点运算量降至2.6×10^9。相较于YOLO 11n mAP50提升12.4个百分点，参数量减少21.3%，浮点运算量降低49.0%;与YOLO v8-Corn、YOLO 11-LAMP、SCD-YOLO 11等典型方法相比，均表现出显著优势，可为资源受限边缘端水稻病虫害实时检测提供技术支撑。

Abstract:

Aiming to address the challenge of deploying deep learning models with large parameter sizes and high computational costs for rice pest and disease detection on resource-constrained agricultural edge devices, a lightweight method, YOLO 11-DG was proposed. With YOLO 11n as the baseline, an FC-C3K2 module incorporating a FlexConv dynamic channel matching mechanism was designed to resolve the channel dimension mismatch that occurred with the original C3K2 during structured pruning. A bounding box regression loss, YIoU, containing a normalized distance penalty term was introduced to alleviate the gradient vanishing problem of CIoU when predicted and ground-truth boxes were highly overlapped. On this basis, an FC-DepGraph structured pruning framework was constructed to automatically align channels across modules during pruning, significantly improving pruning efficiency and model compression. Experimental results showed that YOLO 11-DG achieved an mAP50 of 84.5%, with only 2.36×10^6 parameters and 2.6×10^9 FLOPs. Compared with the baseline YOLO 11n, the mAP50 was increased by 12.4 percentage points, the number of parameters was reduced by 21.3%, and the computational cost dropped by 49.0%. The method also outperformed YOLO v8-Corn, YOLO 11-LAMP, and SCD-YOLO 11, providing an effective lightweight solution for real-time rice pest and disease detection on edge devices.