Infrared Small-Target Segmentation Framework Based on Morphological Attention and Energy Core Loss

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Infrared small-target segmentation (IRSTS) is vital for applications like maritime surveillance and traffic monitoring.
• Existing deep learning models face challenges with dynamic scale variations in infrared small targets, leading to poor localization and missed detections.

Purpose of the Study:

• To develop an advanced infrared small-target segmentation framework to overcome limitations of current methods.
• To enhance the accuracy and robustness of small target detection in infrared imagery.

Main Methods:

• Proposed IRSTS_Unet framework integrating a Dynamic Shape-adaptive Deformable Attention Module (DSDAM) for parameterized feature extraction.
• DSDAM is embedded in feature extraction and cross-layer fusion to focus on target cores and suppress background clutter.
• Introduced a Core Energy-aware Core-Priority loss (CECP-Loss) function to mitigate the 'core dilution' issue in small target segmentation.

Main Results:

• The IRSTS_Unet framework demonstrated superior performance compared to state-of-the-art methods on multiple public datasets.
• Achieved significant improvements in detection accuracy and robustness for infrared small targets.
• Effectively addressed challenges related to dynamic scale variations and core localization.

Conclusions:

• The proposed IRSTS_Unet framework offers a robust solution for infrared small-target segmentation.
• The DSDAM and CECP-Loss contribute to enhanced target core localization and reduced false detections.
• This work advances the field of infrared target detection for critical applications.