Infrared Small Target Detection Algorithm Based on Improved Dense Nested U-Net Network

Area of Science:

• Computer Vision
• Machine Learning
• Infrared Imaging

Background:

• Infrared small target detection is vital for early warning, medical diagnostics, and anti-UAV systems.
• Convolutional Neural Network (CNN) methods struggle with small infrared targets due to feature loss from downsampling.

Purpose of the Study:

• To address limitations of existing methods in infrared small target detection.
• To propose an improved U-Net based model for enhanced small target detection.

Main Methods:

• Developed an improved dense nesting and attention infrared small target detection method based on U-Net (IDNA-UNet).
• Introduced a dense nested interaction module (DNIM) for level-by-level feature fusion and retention of small target features.
• Implemented a bottom-up feature pyramid fusion module for integrating low-level and high-level features.
• Utilized a scale and position sensitive (SLS) loss function for accurate localization and scale differentiation.

Main Results:

• IDNA-UNet effectively incorporates and exploits contextual information of small targets through repetitive fusion and enhancement.
• The proposed method demonstrates significant advantages in intersection over union (IoU), detection probability (Pd), and false alarm rate (Fa).

Conclusions:

• IDNA-UNet offers superior performance for infrared small target detection compared to existing methods.
• The model's architecture and loss function contribute to improved accuracy and reliability in challenging detection scenarios.