摘要
针对山区分散农田病害识别中环境复杂、目标尺度多变及边缘设备算力受限等导致的检测精度低与部署困难等问题,提出一种轻量化MobileNet-YOLOv7融合模型。其核心创新在于:将YOLOv7的ELAN主干替换为改进型MobileNet结构,并引入基于L1范数的动态通道剪枝机制以实现参数压缩;在第3~5阶段中嵌入轻量多尺度注意力模块,结合并行深度可分离卷积与通道激励生成联合注意力权重;同时重构检测头并采用跨阶段部分特征融合策略,实现浅层细节与高层语义信息的逐级融合。模型通过两阶段训练与TensorRT INT8量化,适配Jetson Nano等边缘设备部署。实验结果显示,该方法在多尺度病害目标检测中的AP@0.5均值为0.764~0.891,AP@0.75均值为0.621~0.842;在复杂背景下mAP@0.5∶0.95达71.8%~73.2%,Precision@0.5为84.9%~86.1%;在Jetson Nano上平均功耗为(4.82±0.11)W,温升控制在(18.3±0.7)℃。与现有轻量化模型相比,本方法在精度、效率与稳定性方面均具有明显优势,为山区农业智能感知提供了更具适应性的技术路径。
To address the challenges of low detection accuracy and difficult deployment in identifying diseases in scattered mountainous farmlands caused by complex environments,variable target scales,and limited computing power on edge devices:this study proposes a lightweight MobileNet-YOLOv7 fusion model.The core innovations of this research include:replacing the ELAN backbone of YOLOv7 with an improved MobileNet architecture and introducing a dynamic channel pruning mechanism based on the L1 norm to achieve parameter compression;embedding a lightweight multi-scale attention module from stage3 to stage5,which combines parallel depthwise separable convolutions with channel excitation to generate joint attention weights;and reconstructing the detection head while adopting a cross-stage partial feature fusion strategy to progressively integrate low-level details with high-level sermantic information.The model undergoes two-stage training and TensorRT INT8 quantization to enable deployment on edge devices such as the Jetson Nano.Experimental results show that the proposed method achieves an AP@0.5 range of 0.764-0.891 and an AP@0.75 range of 0.621-0.842 for multi-scale disease target detection.Under complex backgrounds,the mAP@0.5:0.95 reaches 71.8%-73.2%,with a Precision@0.5 of 84.9%-86.1%.On the Jetson Nano,the average power consumption is(4.82±0.11)W,and the temperature rise is controlled within(18.3±0.7)℃.Compared to existing lightweight models,this approach demonstrates significant advantages in accuracy,efficiency,and stability,offering a more adaptable technical pathway for intelligent agricultural perception in mountainous regions.
作者
万小雨
张信得
李绍稳
WAN Xiaoyu;ZHANG Xinde;LI Shaowen(School of Mathematics and Statistics,Hefei Normal University,Hefei 230601,China;Institute of Industrial Crops,Anhui Academy of Agricultural Sciences,Hefei 230001,China;School of Information Science and Technology,Anhui Agricultural University,Hefei 230001,China)
出处
《湖州师范学院学报》
2026年第2期45-58,共14页
Journal of Huzhou University
基金
合肥师范学院校级科研项目(2024KY63)。
