基于几何表征学习的弱监督旋转目标检测

打开文本图片集

中图分类号：TP391 文献标志码：A 文章编号：1671-8755（2025）02-0094-11

Abstract： To address the challnges of high complexity and annotation costs in general rotated object detection for remote sensing images， this paper proposed a weakly supervised rotated object detection model based on geometric representation learning.The proposed method utilized only horizontal bounding box annotations for training and employed a dual-branch architecture with shared backbone and neck networks.The weakly supervised branch learned the position，aspect ratio，and scale consistency of rotated bounding boxes from horizontalannotations，while theself-supervised branch enhanced rotation consistency.To improve feature representation and contextual interaction，the model introduced a shalow feature enhancement module and proposed a geometric vector representation for rotated bounding boxes， thereby improving the accuracy of rotation consistency learning. For bounding box regression，a distance loss based on vertex coordinates （FPD Loss）was introduced to reduce the sensitivity of size regression to angle precision. Experimental results on the public remote sensing datasets DOTA and DIOR -R demonstrate that the proposed model achieves accuracies of 79.33% and 58.50% ，respectively，outperforming the H2RBox algorithm by 4.8 and 1.5 percentage points. The proposed method improves accuracy while reducing computational cost and complexity under the condition of horizontal bounding box annotations， providing a novel solution for rotated object detection in remote sensing images.

Keywords： Remote sensing image;Weakly supervised learning；Rotated object detection; Vector representation；Feature enhancement

随着遥感图像及相关技术的快速发展，目标检测已作为遥感应用中的核心技术之一，广泛应用于环境监测[1]、城市规划[2]、灾害响应[3]等领域。（剩余15385字）