摘要
针对玉米雄穗检测中速度较慢的问题,提出一种基于YOLOv5s改进的轻量化快速检测算法——PSOS-YOLOv5s。所提算法在主干网络中采用轻量级PP-LCNet替换CSPDarknet53,采用2种不同的深度可分离卷积与注意力机制的组合来构成基本块,降低模型复杂度并加快检测速度;在预测部分采用SimOTA标签匹配策略替换YOLOv5s中的标签匹配策略,采用中心先验思想获得精准的先验知识,提出动态k策略过滤冗余标签,提高模型对正样本的快速选取能力;在预测部分采用SIOU Loss替换GIOU Loss,引入角度损失因子来降低回归自由度、加快收敛速度、节省训练时间,重新定义惩罚指标,提高检测精度。实验结果表明,在玉米雄穗数据集中,提出的改进算法相比于YOLOv5s,模型参数量降低52.86%,模型的检测精度提升0.6%,模型的检测速度提升65.5%。改进后的算法提升效果明显,可以满足大规模玉米雄穗快速检测的要求。
In order to solve the problem of slow speed in maize tassel detection,a lightweight and fast detection algorithm PSOS-YOLOv5s based on YOLOv5s improvement is proposed.The proposed algorithm uses lightweight PP-LCNet to replace CSPDarknet53 in the backbone network,and uses a combination of two different depth-separable convolutions and attention mechanisms to form basic blocks,reducing model complexity and speeding up detection;in the prediction part,the SimOTA label matching strategy is used to replace the label matching strategy in YOLOv5s,the central prior idea is used to obtain accurate prior knowledge,and the dynamic k strategy is proposed to filter redundant labels to improve the model s ability to quickly select positive samples;SIOU loss is used in the prediction part to replace GIOU Loss and the angle loss factor is introduced to reduce the degree of freedom of regression,speed up the convergence speed,save training time,and redefine the penalty index to improve detection accuracy.Experimental results show that in the maize tassel data set,compared with YOLOv5s,the number of model parameters of the proposed improved algorithm is reduced by 52.86%,the detection accuracy of the model is increased by 0.6%,and the detection speed of the model is increased by 65.5%.It shows that the improved algorithm has obvious improvement effect and can meet the requirements of rapid detection of large-scale maize tassels.
作者
胡阵
马宗军
黄传宝
赵景波
唐勇伟
郝凤琦
HU Zhen;MA Zongjun;HUANG Chuanbao;ZHAO Jingbo;TANG Yongwei;HAO Fengqi(School of Information and Control Engineering,Qingdao University of Technology,Qingdao 266520,China;Shandong Computer Science Center(National Supercomputer Center in Jinan),Qilu University of Technology(Shandong Academy of Sciences),Jinan 250014,China;Key Laboratory of Computing Power Network and Information Security of Ministry of Education,Qilu University of Technology(Shandong Academy of Sciences),Jinan 250014,China)
出处
《无线电工程》
2024年第6期1446-1453,共8页
Radio Engineering
基金
山东省科技型中小企业创新能力提升工程项目(2023TSGC0111,2023TSGC0587)
青岛市民生计划(22-3-7-xdny-18-nsh)
山东省重点研发计划(软科学项目)(2023RZA02017)
山东省重大科技创新工程项目(2019JZZY020603)。
