Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (24): 155-164.DOI: 10.3778/j.issn.1002-8331.2207-0077

• Graphics and Image Processing • Previous Articles     Next Articles

Adaptive Feature Refinement for Oriented Object Detection in Remote Sensing Images

LIU Enhai, XU Jiayin, LI Yan, FAN Shiyan   

  1. 1.School of Artificial Intelligence, Hebei University of Technology, Tianjin 300401, China
    2.Hebei Province Key Laboratory of Big Data Calculation, Hebei University of Technology, Tianjin 300401, China
  • Online:2023-12-15 Published:2023-12-15

自适应特征细化的遥感图像有向目标检测

刘恩海,许佳音,李妍,樊世燕   

  1. 1.河北工业大学 人工智能与数据科学学院,天津 300401
    2.河北工业大学 河北省大数据计算重点实验室,天津 300401

Abstract: Object detection is an important and challenging task in remote sensing research. Remote sensing images are mostly taken from a top-down perspective. Due to their complex background and arbitrary orientation, object detection algorithms in natural scenes face some challenges when directly applied to remote sensing. Aiming at the above problems, this paper proposes an adaptive feature refinement network AFR-Net to generate directed candidate boxes with high matching degree to objects. Firstly, the feature enhancement module is designed to increase the feature representation with discriminative power, so as to improve the ability of capturing spatial details in complex background. Secondly, in order to obtain the directed candidate box adapted to the object direction, an adaptive feature alignment module is proposed to alleviate the spatial misalignment problem between convolution feature and directed objects, and the rotation-invariant feature is obtained. Finally, the rotation sensitive features are obtained by decoupling detection head module and accurate bounding box regression is refined. The proposed network achieves 66.71% and 97.12% accuracy in the publicly available remote sensing object detection datasets DIOR-R and HRSC2016, which are 2.3 and 0.9 percentage points higher than the original algorithm, respectively. At the same time, compared with some mainstream object detection algorithms, the proposed network has certain advantages.

Key words: convolutional neural network, remote sensing image, object detection, feature refinement, feature alignment

摘要: 目标检测是遥感研究中重要且具有挑战性的任务。遥感图像大多通过俯视视角拍摄,其背景复杂、方向任意的特点,使得自然场景中的目标检测算法直接应用于遥感领域会面临一些挑战。针对上述问题,提出一种自适应特征细化网络AFR-Net,生成与物体之间具有高匹配度的有向候选框。设计特征增强模块增加具有判别力的特征表示,提升复杂背景下空间细节的捕捉能力;为得到适应于物体方向的有向候选框,提出自适应特征对齐模块缓解卷积特征与有向目标的空间错位问题,得到旋转不变特征;通过解耦检测头模块获取旋转敏感特征并细化精确的边界盒回归。提出的网络在公开的遥感目标检测数据集DIOR-R和HRSC2016达到了66.71%和97.12%的准确率,相比原始算法分别提高了2.3和0.9个百分点的检测精度,同时与一些主流的目标检测算法相比,该算法具有一定的优越性。

关键词: 卷积神经网络, 遥感图像, 目标检测, 特征细化, 特征对齐