摘要
农田杂草是影响农作物生长的主要因素之一,农田杂草的有效防治与农作物产量息息相关。复杂田间环境下,精准识别玉米秧苗与农田杂草能够指导除草装备作业更加经济和高效。为提高农田目标识别精度和效率,文章基于深度学习技术的目标检测方法,首先使用多苗期、多时段和单一拍摄角度的图像采集方式并配合数据增强方法制作一个特征丰富的数据集。通过减少YOLOv4网络的输出张量为13×13和52×52两个尺度匹配玉米苗和杂草,并用制作数据集作网络训练。训练结果表明,改进后YOLOv4网络训练得到的检测模型在综合性能上优于YOLO v3、原本YOLO v4和主干网络为VGG19的Faster R-CNN;其F_(1)值为0.828,较修改前提升0.031,检测时间缩短0.014s。此外,根据试验可知数据量和数据增强方式均对模型产生不同程度影响;不同类别的目标进行单一训练比多类别目标组合训练得到检测效果更好。
Farmland weed is one of the main factors affecting the growth of crops,the effective control of farmland weed is closely related to crop yields.In a complex field environment,accurate identification of corn seedlings and farmland weeds can guide the operation of weeding equipment to be more economical and efficient.In order to improve the accuracy and efficiency of farmland target recognition,a target detection method was adopted based on deep learning technology.Firstly,a feature-rich data set was created by using the image acquisition method of multiple seedling stages,multiple periods and a single shooting angle,combined with the data enhancement method.The output tensor of the YOLO v4 network was reduced to 13×13 and 52×52 to match the corn seedlings and weeds,and the data set was used for network training.The training results showed that the detection model obtained by the modified YOLO v4 network training was better than that of YOLO v3 in overall performance,the original YOLO v4 and the Faster R-CNN whose backbone network was VGG19;F_(1) value was 0.828,which was an improvement of 0.031.The detection time was shortened by 0.014 s.In addition,according to the experiment,it could be known that the amount of data and the data enhancement method had different degrees of impact on the model;single training for different types of targets was better than multi-category combined training.
作者
权龙哲
夏福霖
姜伟
李海龙
李恒达
娄朝霞
李传文
QUAN Longzhe;XIA Fulin;JIANG Wei;LI Hailong;LI Hengda;LOU Zhaoxia;LI Chuanwen(School of Engineering,Northeast Agricultural University,Harbin 150030,China)
出处
《东北农业大学学报》
CAS
CSCD
北大核心
2021年第7期89-98,共10页
Journal of Northeast Agricultural University
基金
国家自然科学基金项目(52075092)
黑龙江省博士后科研启动基金项目(LBH-Q19007)。
