基于模型微调的空中无人机小样本目标识别方法 |
| |
引用本文: | 黄灿.基于模型微调的空中无人机小样本目标识别方法[J].计算机测量与控制,2024,32(1):268-274. |
| |
作者姓名: | 黄灿 |
| |
作者单位: | 中国人民解放军91550部队 |
| |
摘 要: | 空中无人机目标识别是现代军事、航空领域的迫切需求,由于目前无人机的功能和种类繁多,对于新机型很难采集大量的无人机样本用于训练目标识别模型;针对该问题,提出了一种基于模型微调的空中无人机小样本目标识别方法;方法以Faster R-CNN为基础架构,首先采用具有大量标记样本的常见机型数据预训练Faster R-CNN模型;然后将基础架构最后的分类层替换为余弦度量,构建联合新机型与常见机型的小样本平衡数据集以较小的学习率微调分类层。实验结果表明,在标记样本数量为5、10和50的情况下,基于模型微调的小样本目标识别模型的mAP分别为88.6%,89.2%和90.8%,能够满足空中无人机小样本目标识别任务需求,且优于其它小样本目标识别方法。
|
关 键 词: | 无人机 目标识别 Faster R-CNN 小样本学习 模型微调 |
收稿时间: | 2023/7/28 0:00:00 |
修稿时间: | 2023/8/21 0:00:00 |
Few-shot Object Recognition Method of Aerial Drone Based on Model Fine-tuning |
| |
Abstract: | Aerial drone object recognition is an urgent demand in modern military and aviation fields. Due to the various functions and types of drones at present, it is difficult to collect a large number of new drone samples for training object recognition models. In order to solve this problem, a few-shot object recognition method of aerial drone based on model fine-tuning is proposed. The method is based on the Faster R-CNN architecture, the Faster R-CNN is pretrained by using common drone data with a large number of labeled samples. Then, the last classification layer of the infrastructure is replaced by cosine measurement, and a small balanced dataset combining new and common drones is constructed to fine-tune the classification layer with a small learning rate. Experimental results show that when the number of labeled samples is 5, 10 and 50, the mAP of few-shot object recognition model based on model fine-tuning is 88.6%, 89.2% and 90.8% respectively, which can meet the requirements of few-shot target recognition task of aerial drone, and is better than other few-shot target recognition methods. |
| |
Keywords: | drone object recognition Faster R-CNN few-shot learning model fine-tuning |
|
| 点击此处可从《计算机测量与控制》浏览原始摘要信息 |
|
点击此处可从《计算机测量与控制》下载全文 |
|