Abstract
Infrared small target detection (ISTD) is crucial for civilian and military applications, but accurately identifying weak targets amid complex backgrounds remains challenging. To address this, we propose MSFANet, a multi-scale frequency-and spatial-domain attention network. MSFANet employs a multi-scale mixed attention module for effective feature extraction and fusion, and introduces a dynamic frequency-spatial module (DFSM), which specifically enhances high-frequency signals by Fourier transform of high-level features, preserves details by combining spatial convolution of low-level features, and dynamically fuses them to suppress background noise. Deep supervision and a combined loss strategy further improve segmentation accuracy and robustness. Experiments show MSFANet outperforms state-of-the-art (SOTA) methods on challenging infrared datasets.
Funding Information
  • Natural Science Foundation of Chongqing Municipality (CSTB2025NSCQ-GPX0812)
  • Natural Science Foundation of Chongqing Municipality (CSTB 2025NSCQ-GPX0740)
  • National Natural Science Foundation of China (61675036)
  • Chongqing Municipal Education Commission (KJQN202400814)
  • Chongqing Municipal Education Commission (CY250801)

This publication has 23 references indexed in Scilit: