沈阳农业大学学报2025，Vol.56Issue(6)：68-82,15.DOI:10.3969/j.issn.1000-1700.2025.06.007
水稻叶部病害轻量化检测方法研究
Research on Lightweight Detection Method of Rice Leaf Diseases
摘要
Abstract
[Objective]In order to solve the problems of small target feature extraction difficulty,low detection accuracy in field environment,and poor timeliness of edge device deployment in rice spot detection,the study proposes a lightweight rice leaf disease detection model YOLOv7-TMRTM-prune based on the channel pruning method.[Methods]Firstly,on the framework of the YOLOv7-tiny baseline model,we adopted the MobileNetV3 to replace the backbone network,RCS-OSA to replace ELAN-1,and the introduction of TSCODE to optimize the detection head to enhance the model's ability to extract features of different sizes of spots,and then construct the leaf disease detection model YOLOv7-TMRTM.Then,we performed channel pruning on the spot detection model,and sparse training based on the L1 regularization criterion.The sparsity of 0.008 was selected for sparse training based on the distribution of the scaling factor in the BN layer;a pruning rate of 75%was selected for channel pruning of the model under the constraints of the balance of the three parameters:detection accuracy,speed and model complexity.[Results]The experimental results show that the mAP of YOLOv7-TMRTM-prune is 97.1%,the Params are reduced by 76.4%,the model size is reduced by 77.4%,and the FPS is improved by 43%.The constructed lightweight leaf disease detection model YOLOv7-TMRTM-prune has a precision of 94.9%,an average recall of 93.3%,and an average mAP of 97.1%for detection of rice blast,rice bacterial,and rice brown spot.[Conclusion]This lightweight model not only ensures high detection accuracy,but also significantly reduces the number of model parameters and size,improves detection speed,effectively solves the problems of difficult extraction of small target lesions and poor timeliness of edge device deployment.It provides an efficient technical solution for real-time field detection of rice leaf diseases and can support the needs of early disease prevention and control in smart agriculture.关键词
水稻叶部病害检测/YOLOv7/通道修剪/L1正则化准则/BN层缩放因子/稀疏训练Key words
rice leaf disease detection/YOLOv7/channel pruning/L1 regularization criterion/BN layer scaling factor/sparse training分类
农业科技引用本文复制引用
李晓辉,刘星呈,李静,陈思诺..水稻叶部病害轻量化检测方法研究[J].沈阳农业大学学报,2025,56(6):68-82,15.基金项目
辽宁省教育厅重点攻关项目(JYTZD2023123) （JYTZD2023123）
