基于YOLOX-Tiny有偏特征融合网络的遥感目标检测
Object Detection in Remote Sensing Images based on YOLOX-Tiny Biased Feature Fusion Network
胡昭华 1李昱辉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;南京信息工程大学 江苏省大气环境与装备技术协同创新中心,江苏 南京 210044
- 2. 南京信息工程大学 电子与信息工程学院,江苏 南京 210044
- 折叠
摘要
遥感目标检测在环境监测、电路巡检等领域具有重要意义.然而,遥感图像中存在目标尺度差异大、小目标数量较多、类间相似性与类内多样性较高等难题,导致检测检测精度较低.为了解决上述问题,提出了一种基于YOLOX-Tiny的遥感目标检测模型.首先通过改进多尺度特征融合网络以充分利用浅层细节信息和深层语义信息,提高对小目标的检测能力;其次在预测端引入可形变卷积,提高模型对不同尺度、形状目标的鲁棒性;最后采用SIoU损失函数以让预测框向正确的方向移动,进一步提高模型的定位精度.在遥感数据集DIOR和RSOD进行实验,实验结果表明:在不增加参数量的情况下,改进后的模型分别取得了73.68%和97.12%的检测精度,与其他一些先进模型相比,具有较高的精确度,对重叠目标识别率高,且具有良好的实时性.
Abstract
Remote sensing target detection is of great significance in fields such as environmental monitoring and circuit inspection.However,there are challenges in remote sensing images,such as large differences in target scale,a large number of small targets,high inter class similarity and intra class diversity,which lead to low de-tection accuracy.To solve the above problems,a remote sensing target detection model based on YOLOX-Ti-ny is proposed.Firstly,by improving the multi-scale feature fusion network to fully utilize shallow detail infor-mation and deep semantic information,the detection ability for small targets is enhanced;Secondly,deformable convolution is introduced at the prediction end to improve the robustness of the model to targets of different scales and shapes;Finally,the SIoU loss function is used to move the prediction box in the correct direction,further improving the positioning accuracy of the model.Experiments are conducted on remote sensing datasets DIOR and RSOD,and the experimental results show that without increasing the number of parameters,the im-proved model achieves a detection accuracy of 73.68% and 97.12%,respectively,which is high compared to some other state-of-the-art models,with a high recognition rate of overlapping targets and good real-time per-formance.
关键词
小目标检测/遥感图像/YOLOX/可形变卷积Key words
Small target detection/Remote sensing images/YOLOX/Deformable convolution引用本文复制引用
基金项目
国家自然科学基金项目(61601230)
江苏省自然科学基金项目(BK20141004)
出版年
2024
NETL
NSTL
万方数据