GD-RiceNet：轻量级水稻叶病检测网络

GD-RiceNet: A Lightweight Network for Rice Leaf Disease Detection

摘要: 针对复杂田间环境中水稻叶部病害目标尺度差异大、背景干扰严重且资源受限等问题，本文借鉴水稻叶脉的重复纹理特性，构建轻量级检测网络GD-RiceNet。叶脉纹理的局部自相似性与Ghost模块的轻量化思路相契合，据此，该网络以GhostHGNetV2为骨干减少标准卷积计算，颈部引入Slim-neck实现低开销多尺度特征融合，同时采用DySample动态上采样与DyHead动态检测头缓解轻量化带来的空间细节损失与特征选择不足。在公开田间水稻病害数据集上，GD-RiceNet的mAP@50和mAP@50:95分别达57.3%和35.3%，参数量6.56 M，计算量12.7 GFLOPs，推理速度169 FPS。相比YOLO11s，参数量和计算量分别降低30.4%和40.9%，mAP@50和mAP@50:95分别提升0.6和1.4个百分点，推理速度由185 FPS降至169 FPS。结果表明，GD-RiceNet在检测精度与模型效率之间取得了较好平衡，具备面向农业边缘设备部署的潜力。

Abstract: Rice leaf disease detection in complex field environments is challenged by large variations in lesion scale, strong background interference, and limited edge-computing resources. The local self-similarity of leaf vein textures aligns with the lightweight design principle of the Ghost module. Accordingly, the network uses GhostHGNetV2 as its backbone to reduce standard convolutional computation and incorporates Slim-neck into its neck for low-cost multiscale feature fusion. It also employs DySample dynamic upsampling and the DyHead dynamic detection head to mitigate the loss of spatial detail and inadequate feature selection associated with lightweight design. Under a unified data split and shared test protocol, GD-RiceNet achieves 57.3% mAP@50 and 35.3% mAP@50:95 on a public field rice disease dataset, with 6.56 M parameters, 12.7 GFLOPs, and 169 FPS under the shared test conditions. Relative to YOLO11s, GD-RiceNet reduces parameters and computation by 30.4% and 40.9%, respectively, while improving mAP@50 and mAP@50:95 by 0.6 and 1.4 percentage points; throughput decreases from 185 to 169 FPS. These results indicate a favorable accuracy-efficiency trade-off under the present dataset and test conditions. The model therefore warrants further validation on independent datasets and agricultural edge devices.