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

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Infrared (IR) small target detection faces challenges due to low signal-to-noise ratio (SNR) and limited textural details.
• Existing methods struggle to effectively bridge model-driven physical priors and data-driven learning approaches.

Purpose of the Study:

• To propose ISANet, a novel hybrid encoder-decoder framework for improved IR small target detection.
• To integrate physical priors with deep learning for enhanced target discrimination.

Main Methods:

• ISANet employs a multi-dimensional feature extraction (ICG) module to incorporate physical priors.
• A shuffle attention (SA) module is utilized for semantic refinement.
• An adaptive asymmetric context feature fusion (AACF) module integrates shallow details with deep semantics.

Main Results:

• ISANet significantly outperforms state-of-the-art methods in key performance metrics.
• Demonstrated improvements in detection rate (Pd) and reduction in false alarm rate (Fa).
• Effective preservation of target shape in complex background scenarios.

Conclusions:

• ISANet offers a robust and effective solution for IR small target detection.
• The hybrid approach successfully combines physical insights with data-driven learning.
• The framework shows promise for real-world applications requiring reliable target discrimination.