Improved DETR Method for Building Detection in High-resolution Remote Sensing Images
吴奇鸿 1张斌 1段功豪 2郭昶 1王磊1
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 武汉工程大学智能机器人湖北省重点实验室,武汉 430205
2. 复杂系统先进控制与智能自动化湖北省重点实验室,武汉 430079
折叠
摘要
针对高分辨率遥感影像中建筑目标较小和背景信息冗余带来的挑战,提出了 一种称为FE-DETR(feature enhancement-detection with transformer)的端到端目 标检测 算法.首先,利用 拼接融 合模块(concatenation fusion module,CFM)融合不同尺度的特征层,缓解小建筑目标特征缺失问题;其次,使用全局通道注意力(global channel attention,GCA)模块细化融合后的特征.具体来说,该模块通过构建通道间的关系矩阵,提高模型对目标的感知能力,有效缓解复杂背景信息带来的干扰.最后,在WCH(Wuhan caidian house)、EA(east Asia)和CBC(city building of China)数据集上评估该算法的检测性能.实验结果表明,所提出的改进算法在上述3个数据集上AP50分别提高了 0.8％、0.6％和0.6％,验证了该算法的有效性.
Abstract
Aiming at the challenges brought by small building objects and redundant background information in high-resolution remote sensing images,we propose an end-to-end object detection algorithm called FE-DETR(feature enhancement-detection with transformer).Firstly,the concatenation fusion module(CFM)is used to fuse feature layers of different scales to alleviate the problem of missing small building object features.Secondly,the fused features are refined using a global channel attention(GCA)module.Specifically,this module improves the model's ability to perceive the object by constructing a relationship matrix between channels,and effectively alleviates the interference caused by complex background information.Finally,we evaluate the detection performance of our algorithm on WCH(Wuhan caidian house),EA(east Asia)and CBC(city building of China)datasets.The experimental results show that the improved algorithm proposed in this paper increases the AP50 by 0.8％,0.6％and 0.6％respectively on the above three datasets,which verifies the effectiveness of the improvement in this paper.
关键词
建筑物检测/高分辨率/特征融合/全局通道注意力/DETR
Key words
building detection/high resolution/feature fusion/global channel attention/DETR
维普
万方数据