Self-supervised real-noise fusion infrared small target detection network.

Area of Science:

• Computer Vision
• Machine Learning
• Infrared Imaging

Background:

• Infrared small target detection (IRSTD) is vital for surveillance and monitoring.
• Existing models struggle with generalization and localization due to data scarcity and scale variations.

Purpose of the Study:

• To develop a self-supervised network for improved IRSTD.
• To enhance generalization and localization accuracy in challenging real-world scenarios.

Main Methods:

• A self-supervised real-noise fusion network incorporating authentic background noise.
• Negative sample augmentation for increased dataset diversity and difficulty.
• A composite dynamic scale and location sensitive (DSLS) loss function.
• Self-supervised optimization across multi-scale predictions.

Main Results:

• The proposed method significantly outperforms the leading DNANet on public datasets.
• Achieved IoU improvements of 7.04% and 1.49%, and detection rate increases of 0.91% and 0.96% on two datasets.
• Demonstrated superior generalization and localization capabilities, especially under small-sample conditions.

Conclusions:

• The proposed network effectively addresses data scarcity and target variation issues in IRSTD.
• The real-noise fusion and augmentation strategies enhance model robustness.
• The framework shows strong adaptability and improved performance for infrared small target detection.