基于改进YOLO v5s的垃圾检测算法  

Garbage Detection Algorithm Based on Improved YOLO v5s

作　　者：林梓健 林群煦 张弓 刘成沛 杨智才 张立炜 王晓佳 LIN Zijian;LIN Qunxu;ZHANG Gong;LIU Chengpei;YANG Zhicai;ZHANG Liwei;WANG Xiaojia(School of Railway Transportation,Wuyi University,Jiangmen 529020,China;Guangzhou Institute of Advanced Technology,Guangzhou 511458,China;Luoding Polytechnic,Luoding 527200,China)

机构地区：[1]五邑大学轨道交通学院,广东江门529020 [2]广州先进技术研究所,广州511458 [3]罗定职业技术学院,广东罗定527200

出　　处：《机械工程师》2023年第8期54-57,60,共5页Mechanical Engineer

基　　金：云浮市科技计划项目(2021020401)。

摘　　要：垃圾分拣是一个环境恶劣、重复性高、体力消耗大的岗位,适宜通过智能化设备代替人工进行垃圾分拣。文中提出一种基于YOLO v5s进行改进,用于垃圾识别分类的改进YOLO v5s视觉检测算法。首先进行结构改进,通过改进损失函数、引入K聚类锚框等改进,对2种注意力机制模块及2种嵌入的位置进行比较和选择以提高精度,并通过融合SPPF模块进行提速改进。结构改进后,通过对比实验数种训练策略,进行训练策略改进。同时在搜集到的小型数据集上进行比较,两部分改进后的算法比原算法的m AP提高了1.35%,同时对检测速度影响较小,并与其他算法进行了对比。Garbage sorting is a work with harsh environment,high repeatability and heavy physical consumption.It is suitable to replacing manual garbage sorting with intelligent equipment.Therefore,this paper proposes an improved vision detection algorithm based on YOLO v5s for garbage detection and classification.Firstly,the structure is improved by improving the loss function and fusing the K cluster anchor.Then,two kinds of attention mechanism modules and two kinds of embedded positions are compared and selected to improve the detection accuracy.Meanwhile,the SPPF module is fused to improve the detection speed.After structural improvement,several training strategies are compared and selected to improve the training result.Finally,the comparison on a small data set collected by the author shows that the mAP of the improved YOLOv5s algorithms is improved by 1.35%compared with the original algorithm and it has small effect on the detection speed,and the comparison is made with other algorithms.

关 键 词：YOLO v5s 注意力机制 垃圾检测 视觉检测 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]