Scholarlink

<p>The complex background and objects in arbitrary orientations in remote sensing images present significant challenges for oriented object detection. The existing methods mainly focus on improving oriented bounding box representation, which is still difficult to achieve the shape, texture, and spatial geometric features of oriented objects. To solve this issue, we propose a target pose-guided and feature loss-reduced network (PGLRNet). Our feature acquisition dynamically adjusts the orientation of convolutional kernels based on the pose feature of targets, while increasing the orientation sensitivity of oriented objects in remote sensing scenarios. Furthermore, we also implement an enhanced path feature pyramid network that embeds our specific feature refinement layer to compensate for loss and degradation during feature aggregation. Extensive experiments conducted on the DOTA, DIOR, and HRSC2016 datasets show that our approach achieved state-of-the-art results on the DOTA dataset and competitive performance on the DIOR and HRSC2016 datasets. Code is available at <a href="https://github.com/hylzsyyy/PGLRNet">https://github.com/hylzsyyy/PGLRNet</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

PGLRNet: target pose-guided and feature loss-reduced network for oriented object detection in remote sensing images

  • 期刊论文

  • 发表日期：

    2025年5月30日

查看全文
  • Haiyu Liu,
  • Shuai Zhang,
  • Keyan Ren,
  • Hu Zhao,
  • Xuhong Li,
  • Zhiyu Nie

摘要

The complex background and objects in arbitrary orientations in remote sensing images present significant challenges for oriented object detection. The existing methods mainly focus on improving oriented bounding box representation, which is still difficult to achieve the shape, texture, and spatial geometric features of oriented objects. To solve this issue, we propose a target pose-guided and feature loss-reduced network (PGLRNet). Our feature acquisition dynamically adjusts the orientation of convolutional kernels based on the pose feature of targets, while increasing the orientation sensitivity of oriented objects in remote sensing scenarios. Furthermore, we also implement an enhanced path feature pyramid network that embeds our specific feature refinement layer to compensate for loss and degradation during feature aggregation. Extensive experiments conducted on the DOTA, DIOR, and HRSC2016 datasets show that our approach achieved state-of-the-art results on the DOTA dataset and competitive performance on the DIOR and HRSC2016 datasets. Code is available at https://github.com/hylzsyyy/PGLRNet.