基于旋转去噪Transformer的遥感图像目标端到端检测
End-to-end Remote Sensing Image Target Detection Based on Rotate Denoise Transformer
王建社1
作者信息
- 1. 合肥讯飞数码科技有限公司人工智能研究院,安徽 合肥 230088
- 折叠
摘要
提出一种基于旋转去噪Transformer的遥感图像旋转目标端到端检测方法.首先,针对遥感图像滑窗切片时存在目标切断影响网络学习的问题,提出基于密集栅格的遥感图像数据增强方法,提升网络对不完整目标的预测能力.其次,提出旋转去噪Transformer架构的端到端检测网络,采用PVTv2 提取多尺度特征,通过Transformer编码器对特征进行聚合与增强,利用Transformer解码器对目标解码,通过分类头和回归头直接实现旋转目标预测,即端到端检测.最后,提出旋转对比去噪学习方法,将去噪学习用于旋转目标检测,加速网络收敛.在 2023 全国大数据与计算智能挑战赛数据集上,所提方法的mAP达到了 98.38%.
Abstract
An end-to-end remote sensing image rotation target detection method based on dense grid argumentation and rotate denoise Transformer is proposed.Firstly,in response to the problem of object cutoff during sliding window slic-ing of remote sensing images,which affects network learning,a remote sensing image data argumentation method based on dense grids is proposed to improve the network's prediction ability for incomplete objects.Then,an end-to-end detec-tion network based on the rotate denoise Transformer architecture is proposed,using PVTv2 to extract multi-scale features,and using the Transformer encoder to aggregate and enhance the features.This paper decodes the target by a Trans-former decoder,and directly achieve rotate object prediction through classification and regression heads,which is called end-to-end detection.Lastly,this paper proposes a rotate contrast denoise learning method,which is used for rotate ob-ject detection and accelerates network convergence.The experiment results on the data of the National Big data and Com-putational Intelligence Challenge 2023 shows that the mAP of the proposed method in this paper reached 98.38%.
关键词
遥感图像/旋转目标检测/密集栅格增强/Transformer/端到端检测Key words
remote sensing image/rotate object detection/dense grid argumentation/Transformer/end-to-end detection引用本文复制引用
出版年
2024
NETL
NSTL
万方数据