基于YOLOv7-tiny改进的高性能目标检测器  

An Improved High-Performance Object Detector Based on YOLOv7-tiny

作　　者：朱文旭 史涛 周佳润 刘祖林 刘海鑫 ZHU Wenxu;SHI Tao;ZHOU Jiarun;LIU Zulin;LIU Haixin(College of Electrical Engineering,North China University of Science and Technology,Tangshan 063000,China;School of Electrical Engineering and Automation,Tianjin University of Technology,Tianjin 300000,China)

机构地区：[1]华北理工大学电气工程学院,河北唐山063000 [2]天津理工大学电气工程与自动化学院,天津300000

出　　处：《电光与控制》2024年第9期70-76,共7页Electronics Optics & Control

基　　金：国家自然科学基金(62103298);河北省自然科学基金(F2018209289)。

摘　　要：针对现有YOLO系列目标检测器存在网络参数量过大和检测精度低的问题,提出了一种基于YOLOv7-tiny模型的高性能通用目标检测器YOLOv7-TT。首先,在其Backbone和Neck网络中引入GF-ELAN模块拓展网络的宽度和深度,剔除网络产生的冗余特征,从而减少参数量、降低计算成本;然后,采用改进的SimOTA样本分配方法,优化训练过程中正样本的分配,加快网络收敛速度;最后,使用知识蒸馏方法对模型进行蒸馏训练,使得模型在保证轻量化的同时提高检测精度。实验结果表明,在VOC2007和COCO2017数据集上,YOLOv7-TT相较于YOLOv7-tiny的网络参数量分别减少11%和9.7%,AP分别提高4.2和3.0个百分点,并且利用知识蒸馏的方法,使其模型检测精度进一步提升,在VOC2007数据集上AP达到59.4%,提高了5.3个百分点,有效解决了模型参数量大和检测精度低的问题。Aiming at the problems of large amount of network parameters and low detection accuracy of the existing YOLO-series object detectors,a high-performance universal object detector named YOLOv7-TT is proposed based on YOLOv7-tiny model.Firstly,Generalized and Friendly ELAN(GF-ELAN)module is introduced into Backbone and Neck networks to expand the width and depth of the network and eliminate the redundant features generated by the network,so as to reduce the parameter quantity and computation cost.Then,the improved SimOTA sample allocation method is used to optimize the allocation of positive samples in the training process and accelerate the convergence speed of the network.Finally,the knowledge distillation method is used to distill and train the model to improve its detection accuracy while ensuring lightweight.The experimental results show that:1)Compared with YOLOv7-tiny,YOLOv7-TT reduces the quantity of network parameters by 11% and 9.7%,and improves the AP by 4.2 and 3.0 percentage points respectively on the VOC2007 and COCO2017 datasets;and 2)The model detection accuracy is further improved by using knowledge distillation,the AP reaches 59.4%(with 5.3 percentage points improved)on the VOC2007 dataset,which effectively solves the problems of large quantity parameters and low detection accuracy.

关 键 词：目标检测 YOLOv7 轻量化 COCO VOC 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]