基于RDN-YOLO的自然环境下水稻病害识别模型研究  

Rice Disease Recognition in Natural Environment Based on RDN-YOLO

作　　者：廖娟 刘凯旋 杨玉青 严从宽 张爱芳 朱德泉[1] LIAO Juan;LIU Kaixuan;YANG Yuqing;YAN Congkuan;ZHANG Aifang;ZHU Dequan(School of Engineering,Anhui Agricultural University,Hefei 230036,China;Institute of Plant Protection and Agricultural Product Quality and Safety,Anhui Academy of Agricultural Sciences,Hefei 230031,China)

机构地区：[1]安徽农业大学工学院,合肥230036 [2]安徽省农业科学院植物保护与农产品质量安全研究所,合肥230031

出　　处：《农业机械学报》2024年第8期233-242,共10页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家重点研发计划项目(2022YFD2001801-3);国家自然科学基金项目(32201665)。

摘　　要：针对自然环境下水稻病害识别准确度易受复杂背景干扰、病害类间差异小难以准确识别等问题,以提高水稻病害识别精度并进行模型的有效轻量化为前提,提出了一种水稻病害识别网络模型(RiceDiseaseNet,RDN-YOLO)。以YOLO v5为基本框架,在主干网络的特征提取阶段嵌入跨阶段部分网络融合模块(C2f),增强模型对病害特征的感知能力,并引入空间深度转换卷积(SPDConv),扩展模型的感受野,进一步提升模型对小病斑特征提取能力;在颈部网络嵌入SPDConv结构,并利用轻量级卷积GsConv替换部分标准卷积,提高颈部网络对病害部位的定位和类别信息预测的准确性及推理速度;以穗瘟病、叶瘟病、胡麻斑病、稻曲病和白枯病5种常见水稻病害为研究对象,在自然环境下采集水稻病害图像,制作水稻病害数据集,进行模型训练与测试。实验结果表明,本文模型病害检测精确率高达94.2%,平均精度均值达93.5%,模型参数量为8.1 MB;与YOLO v5、Faster R-CNN、YOLO v7、YOLO v8模型相比,模型参数量略大于YOLO v5,但平均精度均值最高约高12.2个百分点,在一定程度上减轻模型复杂度的同时获得良好的水稻病害识别效果。Rice diseases such as brown spot,white leaf blight,bacterial blight and rice blast occur frequently during rice development stages,causing serious losses in rice production.Aiming at the challenges in accurately identifying rice diseases under natural conditions,where background is complex,and differences between disease classes are subtle,a rice disease detection network model(RDN-YOLO)was proposed to improve the accuracy of rice disease detection.Firstly,the YOLO v5 network was used as the basic framework,and the C2f module was embedded in the backbone network to enhance the model's perception of disease features.Besides,the SPDConv was introduced in the backbone network to expand the model's perception field and further improve the feature extraction ability of minor disease spots.Secondly,the SPDConv was embedded in the neck network,and the lightweight convolution GsConv was used to replace the standard convolution,which can improve the accuracy of positioning of the disease site and prediction of category information and inference speed,contributing to higher accuracy.The model was trained and tested on a dataset comprising images of five common rice diseases:ear blast,leaf blast,leaf spot,smut,and bacterial blight,where the dataset were collected under natural environment.Experimental results showed that the proposed model achieved a disease detection accuracy of 94.2%with mAP of 93.5%and model parameters of 8.1 MB.Compared with other models YOLO v5,Faster R-CNN,YOLO v7 and YOLO v8,the complexity of the proposed model was only slightly lower than that of YOLO v5,but the mAP was approximately 12.2 percentage points than that of YOLO v5,which signified a notable advancement in rice disease detection,achieving high accuracy while reducing model complexity to a certain extent.

关 键 词：水稻病害识别 YOLO v5 跨阶段部分网络融合模块 空间深度转换卷积 轻量化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]