Infrared and visible image fusion using salient decomposition based on a generative adversarial network

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Infrared (IR) and visible image fusion is crucial for enhanced scene understanding.
• Existing fusion methods often struggle to preserve both detailed textures and thermal signatures.

Purpose of the Study:

• To develop an advanced image fusion technique addressing the limitations of current methods.
• To improve the preservation of texture details and thermal information in fused images.

Main Methods:

• A novel fusion method employing local non-subsampled shearlet transform (LNSST) and generative adversarial networks (GANs).
• Decomposition of source images into basic and salient components using LNSST.
• Fusion of components using two GANs, followed by reconstruction via inverse LNSST.

Main Results:

• The proposed method demonstrates superior performance in fusing IR and visible images.
• Quantitative evaluation using eight objective parameters confirms enhanced preservation of texture and thermal details.
• Comparison against ten other methods highlights the state-of-the-art capabilities.

Conclusions:

• The LNSST-GAN based fusion method effectively addresses the challenges in IR and visible image fusion.
• This technique offers significant improvements in maintaining image fidelity and information content.