基于YOLOv4网络的棉花顶芽精准识别方法  

Precise Recognition Method of Cotton Top Buds Based on YOLOv4 Network

作　　者：刘海涛[1] 韩鑫[1] 兰玉彬 伊丽丽 王宝聚 崔立华 LIU Haitao;HAN Xin;LAN Yubin;YI Lili;WANG Baoju;CUI Lihua(School of Agricultural Engineering and Food Science,Shandong University of Technology,Shandong Zibo 255000,China;Shandong Agricultural Aviation Intelligent Equipment Engineering Technology Research Center,Shandong Zibo 255000,China;Shandong Lyufeng Agricultural Group Co.,Ltd.,Shandong Binzhou 256600,China)

机构地区：[1]山东理工大学农业工程与食品科学学院,山东淄博255000 [2]山东省农业航空智能装备工程技术研究中心,山东淄博255000 [3]山东绿风农业集团有限公司,山东滨州256600

出　　处：《中国农业科技导报》2022年第8期99-108,共10页Journal of Agricultural Science and Technology

基　　金：财政部和农业农村部国家现代农业产业技术体系棉花产业体系项目(CARS-15-22);山东省引进顶尖人才“一事一议”专项(鲁政办字[2018]27号);淄博市重点研发计划项目(2019ZBXC053,2019ZBXC143)。

摘　　要：为实现非接触、低成本、精准识别棉花顶芽,提出一种基于YOLOv4网络在复杂环境下对棉花顶芽进行精准识别的方法。利用K-means算法对棉花顶芽数据集进行聚类,优化先验框改善网络检测精度和速度,得到最优权值模型。对聚类前后模型以及与其他目标检测模型在棉花顶芽检测性能上进行了对比试验,并探究了顶芽在逆光和遮挡环境下,不同模型的检测性能。结果表明:该模型在测试集的平均检测精度(AP)、精确率(P)、召回率(R)、调和平均值(F1)比原模型分别提高0.36%、1.73%、0.52%、1.16%,单张图像平均检测时间缩短0.28 s;对比SSD、YOLOv3、Tiny-YOLOV4模型,该模型检测精确率和F1值最高,性能均衡;在自然场景处于逆光状态下,YOLOv4模型检测顶芽效果好于其他模型,且逆光环境对检测影响小;在遮挡条件下各个模型检测精度均有不同程度下降。In order to achieve non-contact,low-cost accurate detection of cotton top buds,this paper proposed a recognition method based on the YOLOv4 network to detect the cotton top buds in a complex environment. The Kmeans algorithm was used to cluster the cotton top buds data set,and the anchors was optimized to improve the precision and speed of network detection,and the optimal model was selected through analysis. Compared with the performance of model before and after optimization,and the optimal weight model was compared with other object detection models in the detection performance of cotton top buds. The results showed that the average precision(AP)of optimal model in the test set,precision(P),recall(R),F1value of the optimize model increased 0.36%,1.73%,0.52%,1.16% compared with the original YOLOv4 model,and the average detection time of a single image shortened 0.28 s;compared with SSD,YOLOv3,Tiny-YOLOV4 models,the optimize model had the highest F1value and recognition precision,and its performance was more balanced. The top buds at a backlight state in natural scenes had no effect on model detection,and the detection accuracy of each model decreased at different degrees under occlusion conditions.

关 键 词：棉花顶芽 YOLOv4 深度学习 K-MEANS 图像识别 

分 类 号：S126[农业科学—农业基础科学] TP391.4[自动化与计算机技术—计算机应用技术]