TNLRS: Target-Aware Non-local Low-Rank Modeling with Saliency Filtering Regularization for Infrared Small Target...

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Infrared small target detection is crucial but challenging.
• Existing local low-rank models overlook inter-patch correlations, limiting accuracy for arbitrary target shapes.
• Texture information and precise target identification are often compromised.

Purpose of the Study:

• To propose a novel target-aware method for infrared small target detection.
• To enhance detection by incorporating non-local patch correlations and saliency learning.
• To develop a robust framework for complex infrared scenes.

Main Methods:

• A non-local low-rank model with saliency filter regularization is proposed.
• Non-local patch construction combines similar patches for better spatial sparsity.
• A saliency filtering regularization term based on entropy encourages target saliency learning.
• The framework is optimized using the alternative direction multiplier method (ADMM).

Main Results:

• The proposed method demonstrates improved robustness in complex infrared scenes.
• It outperforms several state-of-the-art methods in experimental evaluations.
• Joint target saliency learning is enabled in a lower-dimensional manifold.

Conclusions:

• The novel target-aware non-local low-rank model effectively addresses limitations of previous methods.
• The saliency filter regularization preserves local contexts and avoids matrix approximation errors.
• The ADMM-solved framework provides a robust solution for infrared small target detection.