基于改进YOLOv5s网络的实时输液监测  

Real-time infusion monitoring based on improved YOLOv5s network

作　　者：陈镇国 胡国清[1] 付西敏 陈佳 戈明亮[1] CHEN Zhen-guo;HU Guo-qing;FU Xi-min;CHEN Jia;GE Ming-liang(School of Mechanical and Automotive Engineering,South China University of Technology,Guangzhou 510640,China;Research and Development Department,Guangdong Universal Wisdom Medical Technology Limited Company,Guangzhou 510520,China;School of Chinese Medicine,Guangdong Pharmaceutical University,Guangzhou 528458,China)

机构地区：[1]华南理工大学机械与汽车工程学院,广东广州510640 [2]广东寰宇智慧医疗科技有限公司研发部,广东广州510520 [3]广东药科大学中医学院,广东广州528458

出　　处：《计算机工程与设计》2023年第1期269-276,共8页Computer Engineering and Design

基　　金：广东省自然科学基金项目(2016A030313520)。

摘　　要：针对现有输液监测方式无法同时监测输液速度和输液余量的问题,提出一种基于改进YOLOv5s网络的实时输液监测方法。在原有网络的基础上,融合Mixup数据增强,提高网络的泛化能力;以ACON-C作为激活函数,设计一种基于改进EfficientNetV2的轻量化主干网络,用于改善网络模型表达能力;特征融合阶段引入注意力机制,强化小目标液滴特征;通过Cluster-NMS方法对候选框进行分组集群以区分相似的液位和滴管特征。实验结果表明,所提方法与YOLOv5s相比,参数量下降了31%,mAP提升了1%。在复杂输液环境中,能够实现高精度实时监测输液速度和输液余量。A real-time infusion monitoring method based on the improved YOLOv5s network was proposed to address the problem that existing infusion monitoring methods cannot simultaneously monitor the infusion speed and the remainder of infusion.Based on the original network,the Mixup data augmentation method was integrated to improve its generalization ability.A lightweight backbone network based on the improved EfficientNetV2 model was proposed to enhance the expression ability of the model using ACON-C as the activation function.An attention mechanism was included during the fusion stage to improve the properties of small droplet targets.Meanwhile,the bounding boxes were grouped and clustered using the Cluster-NMS method to distinguish similar liquid level and dropper characteristics.Experimental results show that compared with YOLOv5s,the parameters are reduced by 31%and mAP is increased by 1%.High-precision real-time monitoring of infusion speed and the remainder is verified achievable in a complex infusion environment.

关 键 词：输液监测 YOLOv5s 数据增强 激活函数 轻量化网络 注意力机制 分组集群 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]