Cross spectral and spatial scale non-local attention based unsupervised pansharpening network

Pansharpening means fusing the low spatial resolution multispectral image (LRMSI) and panchromatic (PAN) to get high (HRMSI). Due powerful feature learning ability of deep-learning (DL), DL-based unsupervised fusion methods have been developed explosively. However, most are difficult fully explore utilize correct spectral correlation between LRMSI, HRMSI, PAN images. In addition, CNN-dominated framework is limited by its local without exploring global dependency further enhance feature. Therefore, exploit correlations images dependency, we designed a cross-scale network (CSFNet). This composed two cross scale's nonlocal attention blocks effectively fuse LRMSI features. And strategy implemented mapping computed similarity from scale outputs reconstructed HRMSI The experimental results on datasets show that it achieves state-of-the-art performance compared other methods.