IR-WSANet: A lightweight network with frequency-spatial joint optimization for infrared small target detection

Area of Science:

• Computer Vision
• Signal Processing
• Artificial Intelligence

Background:

• Infrared Small Target Detection (IRSTD) is crucial for military and aerospace applications.
• Traditional methods struggle with small targets, noise, and feature sparsity, limiting accuracy and robustness.

Purpose of the Study:

• To enhance the detection performance of infrared small targets.
• To address challenges in accuracy and robustness in complex scenes.

Main Methods:

• Proposed IR-WSANet, a lightweight network based on an improved YOLOv10.
• Introduced discrete wavelet transform convolution (DWaveletConv) for multi-band feature fusion and noise suppression.
• Developed a cooperative module (POS-SHSA) integrating POSConvEmbedding and partial channel single-head self-attention (SHSA) for improved positioning accuracy.

Main Results:

• Achieved high mAP scores: 97.2% on SIDD-City, 82.6% on SIDD-Mountain, and 82.8% on HIT-UAV.
• Outperformed baseline YOLOv10 by 2.8% to 14.1% mAP.
• Improved F1 score by up to 14.8% with low computational cost (27.9 GFLOPs) and real-time performance (42.8 FPS).

Conclusions:

• IR-WSANet significantly enhances infrared small target detection in complex scenes.
• The frequency-spatial joint optimization strategy effectively improves detection accuracy and robustness.
• The model offers a practical solution for real-time IRSTD applications.