基于YOLOv5s的水稻种子逐粒检测装置研制  

Development of Rice Seed Grain-by-grain Detection DeviceBased on YOLOv5s Small Target Detection Technology

作　　者：吴迅[1] 翁伟[1] 李香格 朱同宇 WU Xun;WENG Wei;LI Xiang-ge;ZHU Tong-yu(Intelligent Manufacturing School,Fujian Information Vocational and Technical College,Fuzhou 350013,China;School of Information Engineering,Huzhou Normal University,Huzhou 313000,China)

机构地区：[1]福建信息职业技术学院智能制造学院,福建福州350013 [2]湖州师范学院信息工程学院,浙江湖州313000

出　　处：《兰州石化职业技术大学学报》2024年第3期21-26,共6页Journal of Lanzhou Petrochemical University of Vocational Technology

基　　金：浙江省自然科学基金公益项目(TGN23C130011);国家自然科学基金资助项目(31701512)。

摘　　要：针对农机化智能检测场景下的水稻种子批量单粒活力检测的需求,结合计算机视觉目标检测技术,设计并研制气吸式种子逐粒检测装置。为解决水稻种子这一小目标的检测效果不佳和漏检问题,提出一种基于YOLOv5s的改进算法。在检测网络中设计上下文特征融合机制,利用新的特征融合方式增强网络多尺度空间和通道信息表征,并添加CA注意力模块,在此基础上改善网络特征融合性能。设计由排列装置、传送装置、检测装置、分选装置等组成的水稻种子逐粒分选线。选取中嘉早17、甬优538和中浙优10三种不同外形的水稻为试验对象,以吸嘴直径、转速、压力作为自变量,以吸空率、单粒吸附率和识别正确率为考核指标,对影响水稻种子检测效果进行试验研究和分析,对于三种不同形状的种子分别选取恰当的吸嘴直径、转速、压力能够有效地吸附并逐粒排列。试验结果表明:改进算法结合研制的系统能满足不同场景检测要求。To meet the demand for batch single grain vitality detection of rice seeds in the intelligent detection scenario of agricultural mechanization,an air suction-based seed-by-seed detection device is designed and developed,combining computer vision object detection technology.In order to address the poor detection performance and missed detection issues of rice seeds as small targets,an improved algorithm based on YOLOv5s is proposed.A context feature fusion mechanism is designed in the detection network to enhance the representation of multi-scale spatial and channel information in the network using a new feature fusion method.Additionally,a CA attention module is added to improve the feature fusion performance of the network.A rice seed sorting line consisting of arrangement,conveying,detection,and sorting devices is designed.Three different-shaped rice varieties,Zhongjiazao 17,Yongyou 538,and Zhongzheyou 10,are selected as experimental objects.The suction nozzle diameter,rotation speed,and pressure are chosen as independent variables,while the empty suction rate,single grain adhesion rate,and recognition accuracy are used as evaluation indicators to study and analyze the factors influencing the detection effect of rice seeds.Appropriate suction nozzle diameters,rotation speeds,and pressures are selected for each of the three different-shaped seeds to effectively adsorb and arrange them grain by grain.Experimental results show that the improved algorithm combined with the developed system can meet the requirements of different detection scenarios.

关 键 词：小目标检测 特征融合机制 逐粒排序 气吸式 单粒吸附率 

分 类 号：S339.3[农业科学—作物遗传育种]