Scholarlink

<p>Infrared small target detection plays a crucial role in military surveillance and disaster monitoring. However, challenges remain due to low signal-to-noise ratio, small target size, and complex backgrounds. Existing deep learning methods often suffer from feature dilution and attention drift when extracting weak target features, leading to low detection accuracy and high model complexity, which limits real-time applicability. To address these issues, we propose a lightweight detection network, MDS-DWGANet, inspired by traditional ideas of multi-directional structure modeling and region-selective enhancement. The network incorporates two key modules: the Multi-Directional Sparse Convolution module (MDSC) builds sparse convolution branches in four directions (horizontal, vertical, and two diagonals) to simulate directional gradient responses, and employs a sparse gating mechanism to focus on high-response edge regions, maintaining structural awareness while reducing parameter overhead. The Dynamic Weight Guided Attention module (DWGA) is designed to mitigate feature dilution and attention drift by integrating four attention pathways with a dynamic channel allocation mechanism, adaptively enhancing salient regions and suppressing redundant interference. DWGA can be viewed as a deep generalization of regional contrast enhancement, improving feature selectivity and robustness in complex scenarios. Experiments on NUDT-SIRST and IRSTD-1&#xa0;K datasets show that MDS-DWGANet achieves 98.30% / 91.24% in detection probability (Pd) and 93.04% / 67.04% in mean Intersection over Union (mIoU), with only 1.4&#xa0;M parameters, demonstrating a favorable balance between accuracy and efficiency.</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

MDS-DWGANet: a lightweight network for infrared small target detection via directional sparse convolution and dynamic attention

  • 期刊论文

  • 发表日期：

    2025年11月21日

查看全文
  • Xiang Liu,
  • Mu Qiao,
  • Bo Wang,
  • Qinghong Sheng,
  • Jun Li,
  • Xiao Ling

摘要

Infrared small target detection plays a crucial role in military surveillance and disaster monitoring. However, challenges remain due to low signal-to-noise ratio, small target size, and complex backgrounds. Existing deep learning methods often suffer from feature dilution and attention drift when extracting weak target features, leading to low detection accuracy and high model complexity, which limits real-time applicability. To address these issues, we propose a lightweight detection network, MDS-DWGANet, inspired by traditional ideas of multi-directional structure modeling and region-selective enhancement. The network incorporates two key modules: the Multi-Directional Sparse Convolution module (MDSC) builds sparse convolution branches in four directions (horizontal, vertical, and two diagonals) to simulate directional gradient responses, and employs a sparse gating mechanism to focus on high-response edge regions, maintaining structural awareness while reducing parameter overhead. The Dynamic Weight Guided Attention module (DWGA) is designed to mitigate feature dilution and attention drift by integrating four attention pathways with a dynamic channel allocation mechanism, adaptively enhancing salient regions and suppressing redundant interference. DWGA can be viewed as a deep generalization of regional contrast enhancement, improving feature selectivity and robustness in complex scenarios. Experiments on NUDT-SIRST and IRSTD-1 K datasets show that MDS-DWGANet achieves 98.30% / 91.24% in detection probability (Pd) and 93.04% / 67.04% in mean Intersection over Union (mIoU), with only 1.4 M parameters, demonstrating a favorable balance between accuracy and efficiency.