Efficient super resolution-based detail injection network for multispectral pan-sharpening

Pan-sharpening is a panchromatic (PAN)-guided super-resolution (SR) process, focused on enhancing spatial resolution of low-resolution multi-spectral (LRMS). Existing SR-based pan-sharpening methods face two critical limitations: (1) lack of unified framework adaptable to both pan-sharpening and single image SR (SISR), and (2) inefficiencies in handling low-frequency feature redundancy in both spatial and frequency domain. Based on these, we model features in both spatial and frequency domain and propose a SR-based detail injection network (SDINet) for pan-sharpening. SDINet designs a global-local spatial extraction block (GSEB) for multi-scale feature extraction in spatial domain, accompanying a Sobel-based gated fusion block (SGFB) to non-linearly suppress low-frequency redundancy. Fourier frequency domain-based spatial detail injection block (FDIB) helps to extract supplement spatial information and to fluently convert SDINet for SISR task through input modification alone. Full-resolution and reduced-resolution experiments demonstrate the advantages of SDINet in both pan-sharpening and SISR tasks.