Computational Fluid Dynamic Network for Infrared Small Target Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Fluid Dynamics

Background:

• Infrared small target detection (IRSTD) is crucial for maritime rescue and early warning systems.
• Existing methods often rely on complex backbone networks or specific modules.
• Challenges include harsh weather, low illumination, and long imaging distances.

Purpose of the Study:

• To propose a novel perspective on IRSTD learning through pixel movement.
• To introduce a new network architecture, CFD-Net, inspired by computational fluid dynamics.
• To improve the robustness and performance of IRSTD.

Main Methods:

• Formulating IRSTD learning based on the mechanism of pixel movement.
• Developing a computational fluid dynamic network (CFD-Net) using a unilateral differential residual structure.
• Incorporating a switch-controlled multidirectional treatment tank (SMTT) for dynamic path guidance.

Main Results:

• CFD-Net utilizes a third-order accurate unilateral difference equation for forward pixel flow.
• The SMTT dynamically guides pixel streams for robust target representation.
• Evaluations on IRSTD-1k and SIRST datasets demonstrate superior performance over SOTA methods.

Conclusions:

• CFD-Net offers a novel and effective approach to IRSTD.
• The fluid dynamics-inspired design enhances detection capabilities in challenging environments.
• The proposed method achieves state-of-the-art results in infrared small target detection.