Scholarlink

<p>Small object detection in remote sensing imagery presents significant challenges due to sparse feature representations, strong background interference, and dense object distributions. The main obstacles include insufficient feature characterization for small targets during feature extraction and ineffective fusion strategies that hinder the collaborative expression of multi-scale information, thereby limiting cross-layer integration. To address these issues, this paper proposes an end-to-end small object detection framework for remote sensing imagery, termed DA-Net, which integrates a dual-attention mechanism. The proposed architecture incorporates a Multi-Scale Guided Attention (MSGA) module and a Scale-Aware Feature Fusion (SAFF) module. MSGA enhances small object representation in the feature extraction stage by leveraging multi-scale convolutions to model diverse receptive fields and introducing a spatial-channel joint attention mechanism to guide the networks focus toward key regions. SAFF employs both spatial and channel attention mechanisms, combined with learnable fusion weights, to adaptively model the importance of features across different scales, thereby enabling efficient cross-layer integration. Experiments conducted on the NWPU VHR-10 dataset demonstrate that the proposed method achieves a 3.9% improvement in mAP@50 over the baseline, validating its effectiveness in complex remote sensing scenarios and its strong generalization capability for small object detection.</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

An end-to-end remote sensing small object detection framework with dual attention mechanisms

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Leping Lin,
  • Yunhang Zheng,
  • Ning Ouyang

摘要

Small object detection in remote sensing imagery presents significant challenges due to sparse feature representations, strong background interference, and dense object distributions. The main obstacles include insufficient feature characterization for small targets during feature extraction and ineffective fusion strategies that hinder the collaborative expression of multi-scale information, thereby limiting cross-layer integration. To address these issues, this paper proposes an end-to-end small object detection framework for remote sensing imagery, termed DA-Net, which integrates a dual-attention mechanism. The proposed architecture incorporates a Multi-Scale Guided Attention (MSGA) module and a Scale-Aware Feature Fusion (SAFF) module. MSGA enhances small object representation in the feature extraction stage by leveraging multi-scale convolutions to model diverse receptive fields and introducing a spatial-channel joint attention mechanism to guide the networks focus toward key regions. SAFF employs both spatial and channel attention mechanisms, combined with learnable fusion weights, to adaptively model the importance of features across different scales, thereby enabling efficient cross-layer integration. Experiments conducted on the NWPU VHR-10 dataset demonstrate that the proposed method achieves a 3.9% improvement in mAP@50 over the baseline, validating its effectiveness in complex remote sensing scenarios and its strong generalization capability for small object detection.