基于局部自适应特征加权算法的遥感图像目标检测
Remote Sensing Image Target Detection based on Local Adaptive Feature Weighting Algorithm
周子龙 1周杰 1罗宏 1徐蕾 1邵根富2
作者信息
- 1. 南京信息工程大学 电子与信息工程学院,江苏 南京 210044
- 2. 杭州电子科技大学 通信学院,浙江 杭州 310000
- 折叠
摘要
遥感图像目标检测已成为目标检测领域的重要组成部分.为了解决复杂背景下高分辨率遥感图像中小尺度目标的漏检与误检问题,结合YOLOv5s算法并在检测端提出一种局部自适应特征加权算法,通过学习已有的标注框信息将含有目标特征的前景与背景分离,获得前景中具有关键信息的目标局部特征,并自适应计算各层局部特征的空间尺度和权重.同时在主干部分引入全局注意力机制用于加强通道与空间之间的跨维度特征信息交互能力,增强特征间的关联性,以弥补检测端局部特征的全局信息丢失,从而降低目标的漏检率和误检率.实验表明:该改进算法的准确率和召回率均得到一定的提升,其均值平均检测精度mAP达到 72.33%,相较于传统YOLOv5s算法提升了2.66%.
Abstract
Object detection in remote sensing images has become a vital aspect of the overall object detection do-main.To address the problems of missed detection and false detection of small-scale objects in high-resolution remote sensing images with complex backgrounds,a local adaptive feature weighting algorithm is proposed at the detection stage,combined with the YOLOv5s algorithm.By learning the existing label box information,the foreground containing target features is separated from the background,and the local features of the target with key information in the foreground are obtained.The spatial scale and weight of local features of each layer are calculated adaptively.Meanwhile,a global attention mechanism is proposed to enhance the interaction capability of cross-dimensional feature information between channels and spatial dimensions in the backbone,so as to strengthen the correlation between features and compensate for the loss of global information within local fea-tures at the detection stage,thereby reducing the rates of missed detection and false detection of targets.Experi-mental results show that the improved algorithm achieves certain improvements in precision and recall,with a mean average precision reaching 72.33%,representing an increase of 2.66% compared to the traditional YO-LOv5s algorithm.
关键词
深度学习/遥感图像/目标检测/局部自适应特征加权Key words
Deep learning/Remote sensing image/Target detection/Local adaptive feature weighting引用本文复制引用
基金项目
国家自然科学基金面上项目(61971167)
国家自然科学基金面上项目(62101275)
国家自然科学基金面上项目(62101274)
出版年
2024
NETL
NSTL
万方数据