Direction-Coded Temporal U-Shape Module for Multiframe Infrared Small Target Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Signal Processing

Background:

• Single-frame infrared small target (SIRST) detection struggles with extremely dim targets due to indistinctive spatial features.
• Multiframe infrared small target (MIRST) detection leverages temporal information but faces challenges in extracting motion direction.
• Existing convolutional methods are often insensitive to motion direction, limiting their effectiveness in MIRST.

Purpose of the Study:

• To propose a novel Direction-coded Temporal U-shape Module (DTUM) for enhanced MIRST detection.
• To develop a method for effectively extracting motion information from infrared sequences.
• To address the limitations of current methods in detecting small, dim targets amidst clutter.

Main Methods:

• Developed a motion-to-data mapping to differentiate target and clutter motion based on direction.
• Designed a Direction-coded Convolution Block (DCCB) to encode motion direction into features.
• Integrated the DTUM into existing single-frame networks for MIRST capabilities.
• Created the NUDT-MIRSDT dataset, specifically for multiframe infrared small and dim target detection.

Main Results:

• The proposed DTUM significantly improves the detection of infrared small and dim targets.
• Experimental results on the NUDT-MIRSDT dataset demonstrate state-of-the-art performance.
• The method effectively suppresses false alarms by leveraging temporal motion information.
• The DTUM module shows versatility and can be incorporated into various single-frame networks.

Conclusions:

• The DTUM offers a simple yet effective solution for MIRST detection, particularly for dim targets.
• The developed dataset and metrics provide valuable resources for future research in this area.
• This approach advances the capability to detect small, dim targets in challenging infrared scenarios.