结合迁移学习的无人机自主降落场景识别方法对比研究
Comparative Study of Unmanned Aerial Vehicle Autonomous Landing Scene Recognition Methods Combined With Transfer Learning
曹先革 1杨金玲1
作者信息
- 1. 东华理工大学测绘与空间信息工程学院,江西南昌,330013
- 折叠
摘要
无人机若具备对降落场景的感知能力,可提高其在未知环境中自主降落的安全性.本文基于ResNet18/50/101三种预训练网络模型,对比分析了冻结模型参数和微调模型参数两种迁移方式构建的无人机自主降落场景识别模型的特性.在自建数据集和2种公开数据集上进行实验,结果表明:迁移学习可以快速构建网络模型,能很好地完成目标场景识别任务,解决了训练数据不足、标签数据缺少的难题;微调模型参数的迁移方式训练的模型鲁棒性强、泛化性能好,但时间成本高;为更好地解决目标任务,在计算机硬件资源满足要求的情况下可以通过选择具有足够深度,并与目标任务相关性强的预训练模型以微调模型参数方式进行迁移训练.实验结果为基于迁移学习进行场景识别任务的解决提供了方法、模型、参数设置等方面的参考.
Abstract
If UAV has the perception of landing scene,it can improve the safety of autonomous landing in unknown environ-ment. Based on three pre-trained network models of ResNet18/50/101,this paper compares and analyzes the char-acteristics of UAV autonomous landing scene recognition models constructed by two transfer methods of freezing model parameters and fine-tuning model parameters. The experi-mental results on self-built datasets and two kinds of public da-tasets show that transfer learning can quickly construct net-work models,complete target scene recognition tasks well,and solve the problems of insufficient training data and labeled data;The transfer learning mode of fine-tuning parameters has strong robustness and good generalization performance,but the time cost is high;In order to better solve the target task,when the computer hardware resources meet the require-ments,transfer learning can be carried out by fine-tuning mod-el parameters by selecting a pre-trained model with sufficient depth and strong correlation with the target task. The results of this experiment provide a reference for the solution of scene recognition task based on transfer learning in terms of meth-ods,models,parameter settings,etc.
关键词
迁移学习/无人机/自主降落/场景识别/微调模型参数Key words
transfer learning/unmanned aerial vehicle/autonomous landing/scene recognition/fine-tuning model parameters引用本文复制引用
基金项目
东华理工大学科研基金(DHBK2019193)
东华理工大学科研基金(DHBK2019194)
出版年
2024
NETL
NSTL
万方数据