特征融合和相移编码的遥感图像旋转目标检测  

Rotating Object Detection in Remote Sensing Images via Feature Fusion and Phase Shifting Coder

作　　者：周为 吴涛[1] 吴锡[1] 符颖 ZHOU Wei;WU Tao;WU Xi;FU Ying(School of Computer Science,Chengdu University of Information Technology,Chengdu 610200,China)

机构地区：[1]成都信息工程大学计算机学院,四川成都610200

出　　处：《计算机技术与发展》2025年第4期22-28,共7页Computer Technology and Development

基　　金：四川省科技重点研发计划项目(2023YFG0025)。

摘　　要：针对现有的旋转目标检测模型对目标朝向任意、排列密集的遥感图像检测精度较低的问题,该文提出了一种基于多尺度特征融合和两倍步长相移编码的遥感图像旋转目标检测算法。首先利用不同大小和深度的卷积进行特征提取,有效地捕获上下文特征;然后,通过自上而下和自下而上的特征融合策略以获取更加精细和丰富的语义特征图;同时引入了两倍步长相移编码模块(Double Stepsize Phase-Shifting Coder,DSPSC)对于角度进行编码后再求损失,解决了基准模型存在的边界不连续问题,从而进一步提高了模型的检测精度;最后,在大规模数据集DOTA和HRSC2016上的实验结果表明,平均检测精度分别为72.97%和90.10%,较基准网络分别提升了2.89百分点和2.70百分点。与其它对比方法相比,该方法在精度上取得最好结果,证明了其在遥感图像旋转目标检测上的有效性。The existing rotating target detection models suffer from low inference accuracy when detecting targets with arbitrary orientations and dense arrangements in remote sensing images.We propose a remote sensing image rotation target detection algorithm based on multi-scale feature fusion and double stepsize phase-shifting coder.Firstly,we use convolutional layers of different sizes and depths to extract features,effectively capturing contextual features.Then,a top-down and bottom-up feature fusion strategy is employed to obtain finer and richer semantic feature maps.Meanwhile,we introduce a DSPSC(Double Stepsize Phase-Shifting Coder)module to encode angles and compute losses,addressing the discontinuity issue in the baseline model and further improving the detection accuracy.Finally,experimental results on large-scale datasets DOTA and HRSC2016 demonstrate the average detection accuracy is 72.97%and 90.10%,respectively,which is 2.89 percentage points and 2.70 percentage points higher than that of the baseline network.Compared with other comparative methods,the proposed method exhibits advantages in accuracy,demonstrating its effectiveness in remote sensing image rotation target detection.

关 键 词：遥感图像 目标检测 相移编码 边界不连续 旋转框 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] P2[自动化与计算机技术—控制科学与工程]