Scholarlink

<p>Remote sensing images are characterized by diverse categories and a large number of small objects. While the CenterNet algorithm eliminates the need for bounding box regression when handling small objects in remote sensing images, it tends to focus excessively on the center point while neglecting the localization of neighboring points around the objects. To address these challenges, we propose HolisticNet, a neighborhood regression and dynamically adjustment object detection network. To address the complex features of small objects in remote sensing images, we propose a Global Sampling Fusion Enhanced Attention (GSFEA). GSFEA performs global sampling on multi-scale feature maps using a gated point fusion method and integrates a gradient activation strategy to achieve cross-modal information enhancement and fusion, thereby significantly improving feature representation capabilities. For object localization, we propose the Neighborhood Regression Localization Mechanism (NRLM), which constructs a local neighborhood centered on an initial predicted point and employs Transformer-based regression to perform fine-grained modeling of semantic relationships among detection points, thus enhancing the precision of bounding box localization. Considering the temporal variation characteristics inherent in remote sensing imagery, we further design a Dynamic Regularized Temporal (DRT) loss function that leverages dynamic differentials and higher-order gradient mechanisms to conduct second-order spatiotemporal complexity modeling of both the loss value and its gradient, thereby improving the stability and fitting capability of the model during training. Evaluated on the DIOR-H dataset, HolisticNet achieves an AP of 43.82% and a mAP of 71.39%, demonstrating its effectiveness and strong overall performance. The code is available at <a href="https://github.com/aokseen/HolisticNet">https://github.com/aokseen/HolisticNet</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

HolisticNet: remote sensing small object detection algorithm based on neighborhood regression and dynamic adjustment

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Hongli Li,
  • Yaofeng Ma,
  • Gang Lei

摘要

Remote sensing images are characterized by diverse categories and a large number of small objects. While the CenterNet algorithm eliminates the need for bounding box regression when handling small objects in remote sensing images, it tends to focus excessively on the center point while neglecting the localization of neighboring points around the objects. To address these challenges, we propose HolisticNet, a neighborhood regression and dynamically adjustment object detection network. To address the complex features of small objects in remote sensing images, we propose a Global Sampling Fusion Enhanced Attention (GSFEA). GSFEA performs global sampling on multi-scale feature maps using a gated point fusion method and integrates a gradient activation strategy to achieve cross-modal information enhancement and fusion, thereby significantly improving feature representation capabilities. For object localization, we propose the Neighborhood Regression Localization Mechanism (NRLM), which constructs a local neighborhood centered on an initial predicted point and employs Transformer-based regression to perform fine-grained modeling of semantic relationships among detection points, thus enhancing the precision of bounding box localization. Considering the temporal variation characteristics inherent in remote sensing imagery, we further design a Dynamic Regularized Temporal (DRT) loss function that leverages dynamic differentials and higher-order gradient mechanisms to conduct second-order spatiotemporal complexity modeling of both the loss value and its gradient, thereby improving the stability and fitting capability of the model during training. Evaluated on the DIOR-H dataset, HolisticNet achieves an AP of 43.82% and a mAP of 71.39%, demonstrating its effectiveness and strong overall performance. The code is available at https://github.com/aokseen/HolisticNet.