Polarized image fusion strategy based on multi-scale feature fusion

Area of Science:

• Computer Vision
• Image Processing
• Optics

Background:

• Polarized image fusion combines multi-scale polarization data.
• Existing methods struggle with extracting and preserving complementary information.
• Enhanced texture and intensity details are crucial for image analysis.

Purpose of the Study:

• To propose a novel polarized image fusion strategy using a Dense Generative Adversarial Network (D-GAN).
• To enhance the extraction and preservation of complementary information from multi-scale polarization data.
• To improve image details, such as edges and textures, in fused polarized images.

Main Methods:

• A Dense Generative Adversarial Network (D-GAN) framework is proposed.
• The generator utilizes DenseNet architecture for feature extraction from S0, Degree of Linear Polarization (DoLP), and Angle of Linear Polarization (AoLP).
• Adversarial training refines features, incorporating a polarization self-attention (PSA) mechanism and a novel polarization feature preservation loss combined with multi-scale structural similarity loss.

Main Results:

• The proposed D-GAN method demonstrates superior performance compared to existing techniques.
• Experimental results show enhanced information preservation in fused images.
• The method significantly improves image quality metrics, including edge and texture details.

Conclusions:

• The D-GAN based polarized image fusion strategy effectively enhances complementary information extraction and preservation.
• The integrated PSA mechanism and specialized loss functions contribute to noise reduction and improved feature retention.
• This approach offers a significant advancement in polarized image fusion for improved visual analysis.