Semantic-guided polarization image fusion method based on a dual-discriminator GAN

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Conventional polarization image fusion methods lack robustness and targeting due to unaddressed material-specific polarization properties and manual fusion rules.
• Existing strategies fail to effectively fuse intensity and polarization images, limiting detailed image representation.

Purpose of the Study:

• To develop a novel end-to-end network model for robust and targeted polarization image fusion.
• To improve the fusion process by incorporating semantic information and a specialized discriminator for polarization image quality.

Main Methods:

• Proposed a semantic guided dual discriminator generative adversarial network (SGPF-GAN) for polarization image fusion.
• Introduced a polarization image information quality discriminator (PIQD) block to guide the weighted fusion process.
• Utilized an adversarial game between a generator and dual discriminators to identify semantic targets and their modalities (polarization/intensity).

Main Results:

• The SGPF-GAN demonstrated superior performance in qualitative visual effects and quantitative measures compared to conventional methods.
• The proposed fusion approach significantly enhanced the performance of transparent and camouflaged hidden target detection.
• The method also showed significant improvements in image segmentation tasks.

Conclusions:

• The SGPF-GAN offers a superior solution for polarization image fusion, enabling more detailed and accurate image representation.
• This approach has significant implications for improving detection and segmentation of challenging targets in various applications.
• The semantic guidance and dual discriminator architecture provide a robust framework for complex image fusion tasks.