Polarized image fusion strategy based on multi-scale feature fusion

Polarized image fusion aims to combine multi-scale polarization data to generate images with improved texture and intensity details. To enhance the extraction and preservation of complementary information from source images, we propose a novel fusion strategy based on a Dense Generative Adversarial Network (D-GAN). In this framework, the generator, based on a DenseNet architecture, extracts features from intensity images (S0), Degree of Linear Polarization (DoLP), and Angle of Linear Polarization (AoLP). These features are then iteratively refined by the discriminator through adversarial training, improving image details such as edges and textures. A polarization self-attention (PSA) mechanism is integrated to capture critical polarization information while reducing noise. Furthermore, we introduce a polarization feature preservation loss function to promote information retention across different image scales. This loss is combined with a multi-scale structural similarity loss to enhance feature extraction. Experimental results demonstrate that our method surpasses existing techniques in terms of both information preservation and image quality metrics.