基于改进YOLOv5s的可回收垃圾检测方法  

Detection Method of Recyclable Waste Based on Improved YOLOv5s

作　　者：雷建云[1,2] 邹金林 夏梦 梁钧[1,3] LEI Jian-yun;ZOU Jin-lin;XIA Meng;LIANG Jun(College of Computer Science,South-Central Minzu University,Wuhan Hubei 430074,China;Hubei Provincial Engineering Research Center for Intelligent Management of Manufacturing Enterprise,Wuhan Hubei 430074,China;Hubei Provincial Engineering Research Center of Agricultural Blockchain and Intelligent Management,Wuhan Hubei 430074,China)

机构地区：[1]中南民族大学计算机科学学院,湖北武汉430074 [2]湖北省制造企业智能管理工程技术研究中心,湖北武汉430074 [3]农业区块链与智能管理湖北省工程研究中心,湖北武汉430074

出　　处：《武汉纺织大学学报》2023年第3期56-64,共9页Journal of Wuhan Textile University

基　　金：国家民委中青年英才培养计划(MZR20007);湖北省科技重大专项(2020AEA011);新疆维吾尔自治区区域协同创新专项(科技援疆计划)(2022E02035).

摘　　要：对垃圾进行回收益处颇多,不仅可以节约资源,还有助于自然环境保护。在传统的垃圾回收中,一般会消耗大量的人力和物力,本文基于现有单阶段目标检测算法YOLOv5s再结合注意力机制和RFB感受野模块,提出一种兼顾检测速度与精度的YOLOv5s改进模型,该模型可运用于室内智能垃圾回收机器人或垃圾场处理终端中。首先对RFB模块的结构做出调整并利用注意力机制进行改进,在一定程度上克服了RFB模块引入其他不必要特征信息的缺点;然后在算法中引入改进后的RFB模块,使算法能更好地与不同尺度的垃圾物体相匹配,提高了检测的精度;并根据数据集目标物体的特点重新调整了锚框大小。实验结果表明,YOLOv5s-SERFB在数据集TrashNet-Plus上有良好的表现,最终改进模型的mAP为91.7%,相比于原始的YOLOv5s模型高出2.2%,算法能较好地满足实时检测任务的需要,同时表现出良好的检测效果。Recycling has many benefits.It can save resources,help protect the natural environment.Traditional garbage recovery generally consumes a lot of manpower and material resources.Based on the existing single-stage target detection algorithm YOLOv5s,combined with the attention mechanism and RFB receptive field module,this paper proposes an improved YOLOv5s model that takes into account detection speed and accuracy,this model can be applied to indoor intelligent garbage recycling robot or garbage disposal terminal.Firstly,the structure of RFB module is adjusted and the attention mechanism is used to improve it.To some extent,the shortcoming of introducing other unnecessary feature information in RFB module is overcome.Then,the improved RFB module is introduced into the algorithm,so that the algorithm can better match with garbage objects of different scales.and improve the detection accuracy.The size of the anchor frame is adjusted according to the characteristics of the target object in the data set.The experimental results show that YOLOv5s-SERFB has a good performance on the data set TrashNet-Plus,and the final mAP of the improved model is 91.7%,2.2%higher than that of the original YOLOv5s model.The algorithm can better meet the needs of real-time detection tasks,while showing good detection effect.

关 键 词：垃圾检测 YOLOv5s 感受野 注意力机制 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]