Bi-domain fusion pyramid network for pansharpening with deep anisotropic diffusion

Pansharpening is the process of fusing a low-resolution multi-spectral image with a high-resolution panchromatic (PAN) image to generate a high-resolution multi-spectral (HRMS) image. In this paper, a new bi-domain fusion pyramid network with deep anisotropic diffusion, termed as BFP-Net, is proposed to generate a high-quality HRMS image with accurate spectral distribution as well as reasonable spatial structure. Different from previous deep models that solely rely on the supervision of the HRMS reference image, the bi-directional information flow mechanism of our network effectively enlarges the receptive field and addresses resolution differences between input images. Bi-domain fusion integrates spatial-frequency domain information with encouraging model to learn complementary representations. Furthermore, the introduction of deep anisotropic diffusion adaptively preserves and enhances edge details, thereby enhancing the visual quality and structural consistency of the target image. The loss functions ensure the network's training direction, which enhances its generalization across different datasets and produces more robust and accurate results. Extensive quantitative and qualitative experiments on real datasets demonstrate the superiority of our method over existing methods in terms of performance, showing excellent sharpening quality and spectral consistency. The source code is available at https://github.com/qiwenjjin/BFP-Net.