基于无人机图像和改进CSRNet的田间大豆苗识别计数  

Identification and Counting of Soybean Seedlings In the Field Based on UAV Images and Improved CSRNet

作　　者：朱轩池 程铎灌 崔道涵 冯江[1] Zhu Xuanchi;Cheng Duoguan;Cui Daohan;Feng Jiang(College of Electrical and Information,Northeast Agricultural University,Harbin 150030,China)

机构地区：[1]东北农业大学信息与电气工程学院,哈尔滨150030

出　　处：《农机化研究》2025年第5期145-151,共7页Journal of Agricultural Mechanization Research

基　　金：黑龙江省“揭榜挂帅”科技攻关项目(2021ZXJ05A03)。

摘　　要：为了得到大豆田间苗期的出苗率和种植密度,以方便及时间苗、补苗和评价种子质量的好坏,提出了一种卷积神经网络方法来对大豆田间苗期进行准确计数与密度评估。使用无人机采集大豆苗期图像制作了数据集,并对数据集进行数据增强。针对田间复杂地形环境,为减少背景环境的干扰,对数据集图像进行直方图均衡化和明度提高等预处理。根据采集的大豆图像密集的特点,采用拥挤场景识别网络(Congested Scene Recognition Network,CSRNet)来构建并改进网络模型,改用VGG19作为前端网络,并加入卷积注意力模块(Convolutional Block Attention Module,CBAM)提高特征提取效果,并根据模型的豆苗图像密度图求和得到图像中的大豆苗数。试验结果表明:模型在测试数据集上平均绝对误差(Mean Absolute Error,MAE)和均方根误差(Root Mean Square Error,RMSE)分别达到了2.35、4.69;对比目标检测方法YOLOv5s、Faster R-CNN和其他人群密度检测方法CSRNet、MCNN和SANet,本研究的模型MAE分别低1.27、0.4、0.63、1.38、0.41,RMSE分别低2.43、0.54、1.83、3.83、2.06。所提出方法计数性能优异,可以满足无人机快速获取田间大豆出苗率的需要,为出苗率的获取提供了参考。To obtain the seedling emergence rate and planting density of soybean field seedlings,facilitate timely seedling replenishment and evaluate the quality of seeds,this study proposed a convolutional neural network method for accurate counting and density assessment of soybean field seedlings.A dataset was created using soybean seedling images collected under an unmanned aircraft,and data enhancement was performed on the dataset.Pre-processing such as histogram equalization and brightness enhancement was performed on the dataset images for the complex terrain environment in the field and to reduce the interference of the background environment.Based on the dense characteristics of the collected soybean images,the Congested Scene Recognition Network(CSRNet)is used to build and improve the network model,and VGG19 was used as the front-end network instead,and the Convolutional Block Attention Module(CBAM)was added to improve the feature extraction.The experimental results show that the Mean Absolute Error(MAE)and Root Mean Square Error(RMSE)of the model in this study reached 2.35 and 4.69,respectively.The MAE of the model in this study was 1.27,0.4,0.63,1.38,0.41 lower,and the RMSE was 2.43,0.54,1.83,3.83,2.06 lower,respectively,than the target detection methods YOLOv5s,Faster R-CNN and other crowd density detection methods CSRNet,MCNN and SANet.The counting performance was excellent and can meet the need of UAV for rapid acquisition of soybean seedling rate in the field,which provided a reference for the acquisition of seedling rate.

关 键 词：无人机 苗木计数 CSRNet 机器视觉 大豆 

分 类 号：S126[农业科学—农业基础科学]