A Wavelet-Recalibrated Semi-Supervised Network for Infrared Small Target Detection Under Data Scarcity.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Infrared small target detection faces challenges from small target size, low contrast, and limited annotated data.
• Existing methods struggle with data scarcity and effectively utilizing unlabeled infrared images.

Purpose of the Study:

• To propose a wavelet-recalibrated semi-supervised network (WRSSNet) for improved infrared small target detection.
• To address data scarcity by integrating synthetic data augmentation and semi-supervised learning.
• To enhance feature representation for highlighting weak targets.

Main Methods:

• Developed WRSSNet integrating synthetic data augmentation, feature reconstruction, and semi-supervised learning.
• Utilized an improved CycleGAN for converting visible-light images to pseudo-infrared images, expanding training data.
• Designed a wavelet-enhanced channel recalibration and fusion (WECRF) module with wavelet decomposition and attention mechanisms for multi-scale feature fusion.

Main Results:

• WRSSNet demonstrated superior detection accuracy on NUAA-SIRST and IRSTD-1K datasets.
• Achieved significantly lower false alarm rates compared to state-of-the-art methods.
• Maintained low computational complexity, making it efficient for practical applications.

Conclusions:

• The proposed WRSSNet effectively overcomes challenges in infrared small target detection.
• Synthetic data augmentation and the WECRF module are crucial for enhancing performance.
• WRSSNet offers a promising solution for robust and efficient infrared small target detection under limited supervision.