LRCFuse: Infrared and Visible Image Fusion Based on Low-Rank Representation and Convolutional Sparse Learning

Area of Science:

• Computer Vision
• Signal Processing
• Artificial Intelligence

Background:

• Cross-modal image fusion is crucial for multi-sensor systems, but current methods often lose critical features.
• Insufficient fusion information can degrade the correlation between source and fused images.

Purpose of the Study:

• To propose a new cross-modal image fusion method, LRCFuse, that preserves maximum information from source images.
• To effectively extract feature information for improved image analysis and downstream tasks.

Main Methods:

• Utilizing learned low-rank representation (LLRR) blocks for dimensionality reduction and feature extraction.
• Introducing common feature preservation module (CFPM) blocks based on convolutional sparse coding to recover common features.
• Employing a multi-level optimization strategy with various loss functions (pixel, shallow, mid, deep, Sobel) for feature refinement.

Main Results:

• LRCFuse effectively detects infrared salient targets.
• The method preserves additional details from visible images.
• Evaluations show superior fusion results for subsequent downstream tasks compared to existing methods.

Conclusions:

• LRCFuse offers an effective approach to cross-modal image fusion by preserving crucial information.
• The proposed method enhances feature extraction capabilities in multi-sensor systems.
• This advancement leads to improved performance in various image analysis applications.