基于改进YOLOv7模型的水田复杂环境稻株识别  

Rice Plant Recognition in Complex Paddy Field EnvironmentsBased on Improved YOLOv7 Model

作　　者：陈学深[1] 梁俊 汤存耀 张恩造 陈彦学 党佩娜 齐龙[1] Chen Xueshen;Liang Jun;Tang Cunyao;Zhang Enzao;Chen Yanxue;Dang Peina;Qi Long(College of Engineering,South China Agricultural University,Guangzhou 510642,China)

机构地区：[1]华南农业大学工程学院,广州510642

出　　处：《农机化研究》2025年第7期9-17,共9页Journal of Agricultural Mechanization Research

基　　金：广东省自然基金项目(2021A1515010831);广州市科技计划项目(202206010125);广东省杰出青年基金项目(2019B151502056);国家自然科学基金项目(51575195)。

摘　　要：为实现水田复杂环境的稻株精准识别,提出了一种基于改进YOLOv7模型的稻株识别方法。采用离线和在线双重数据增强,提高模型训练效果、增强泛化能力并缓解过拟合现象。YOLOv7模型中主干特征提取网络替换为GhostNet网络,增强模型自适应特征提取能力和简化模型参数计算量。YOLOv7主干特征提取网络中引入轻量级注意力机制,增强主干特征提取网络的特征提取能力。YOLOv7模型中CIoU损失函数替换为EIoU损失函数,提高模型预测框的回归效果。模型对比表明,改进YOLOv7模型的稻株识别平均精度均值为89.3%,相比YOLOv7、YOLOv5s、YOLOXs、MobilenetV3-YOLOv7模型,分别提高了4.1、7.6、6.5、0.7个百分点。田间试验表明,晴天、阴天、藻萍、杂草环境背景下平均精度均值分别为91.2%、89.1%、87.5%、88.4%。研究结果可为水田复杂环境的稻株精准识别提供切实方法。A rice plant recognition method based on the improved YOLOv7 model was proposed to address the accurate recognition of rice plants in the complex environment of paddy fields.Using offline and online data enhancement methods to improve model training effect,enhance generalization,and mitigate overfitting.The backbone feature extraction network was replaced by GhostNet network in YOLOv7 model to enhance the adaptive feature extraction capability of the model and simplify the computation of model parameters.Introduced lightweight attention mechanism in YOLOv7 backbone feature extraction network to enhance the feature extraction capability of the backbone feature extraction network.The CIoU loss function in the YOLOv7 model was replaced with the EIoU loss function to improve the regression effect of the model prediction frame.The ablation tests showed that the overall mean accuracy of the constructed GhostNet-YOLOv7 model was 89.3%,which was 4.1,7.6,6.5 and 0.7 percentage points improved over the original YOLOv7,YOLOv5s,YOLOXs,and MobilenetV3-YOLOv7 models.The mean accuracy means were 91.2%,89.1%,87.5%,and 88.4%in sunny,cloudy,algae and weed backgrounds,respectively,and the accurate recognition of rice was achieved under different lighting conditions and complex backgrounds.The results of the study can provide a practical method for accurate recognition of rice plants in the complex environment of paddy fields.

关 键 词：机器视觉 识别 深度学习 YOLOv7 GhostNet 水稻 

分 类 号：S985.2[农业科学—捕捞与储运]