Infrared small target detection network based on multi-dimensional and multi-scale feature fusion

Infrared (IR) small target detection is hindered by a low signal-to-noise ratio (SNR) and scarce textural details. To bridge the gap between model-driven priors and data-driven learning, we propose ISANet, a hybrid encoder-decoder framework. ISANet incorporates a multi-dimensional feature extraction (ICG) module to inject physical priors, a shuffle attention (SA) module for semantic refinement, and an adaptive asymmetric context feature fusion (AACF) module to integrate shallow details with deep semantics. Experimental evaluations demonstrate that ISANet significantly outperforms state-of-the-art methods in detection rate (<i>P</i>_<i>d</i>), false alarm rate (<i>F</i>_<i>a</i>), and shape preservation, offering a robust solution for target discrimination in complex backgrounds.