基于改进RetinaNet的轻量化航拍目标检测方法
Lightweight Aerial Object Detection Method Based on Improved RetinaNet
刘砚菊 1王雪梅 1宋建辉 1刘晓阳 1蒲家鹏1
作者信息
- 1. 沈阳理工大学自动化与电气工程学院,沈阳 110159
- 折叠
摘要
针对无人机航拍图像中存在目标较小、尺度不一以及目标检测算法模型参数量较大等问题,提出基于RetinaNet的轻量化目标检测算法模型:RetinaNet-S.采用GhostNetV2作为轻量化骨干,增强目标的特征信息;在颈部采用轻量化模块GSConv和VoV-GSCSP,在尽量保证检测精度的同时减少模型的参数量;最后,改进损失函数为Focal SIoU Loss,进一步提升模型的精准度.在VisDrone2019数据集上的实验表明,RetinaNet-S的参数量仅有5.52 M,检测速度提升了9.55 FPS.
Abstract
Aiming at the problems of smaller targets,different scales and large number of model parameters of target detection algorithm in UAV aerial images,a lightweight target detection algorithm model,Retinanet-S,based on RetinaNet is proposed.Firstly,GhostNetV2 is used as the lightweight back-bone to enhance the characteristic information of the targets.Secondly,while the lightweight modules GSConv and VoV-GSCSP are used in the neck to ensure the detection accuracy and to reduce the number of model parameters.Finally,the Loss function is improved into Focal SIoU Loss to further improve the accuracy of the model.Through the experiment on VisDrone2019 dataset,the parameter amount of RetinaNet-S is only 5.52 M,the detection speed is increased by 9.55 FPS.
关键词
航拍图像/目标检测/RetinaNet/轻量化Key words
aerial images/object detection/RetinaNet/lightweight引用本文复制引用
基金项目
辽宁省教育厅高等学校基本科研项目(LJKZ0275)
沈阳市中青年科技创新人才支持计划项目(RC210247)
辽宁省属本科高校基本科研业务费专项资金资助项目()
出版年
2024
国家科技期刊平台
NSTL
万方数据