Reference-Based Super-Resolution Network for Remote Sensing Image via Position-Constraint

Monitoring ecological and geological environments with high-resolution(HR)images for long time series is useful in remote sensing. However，capturing HR images for long time series is difficult. Therefore，most previous works reconstructed HR images by single image super-resolution(SR) algorithm. The visual results of these methods are smooth due to limited information in single low-resolution(LR)images. Because multiple observations of the same region are typically captured by different satellites at different times，the HR images can be used as references for the target LR images. Thus，we proposed performing reference-based super-resolution network for remote sensing images by imposing constraints on matching positions. First，a position-encoding-based transformer was used to match and transfer the HR reference details to the LR input in the feature space. The position-constraint strategy was implemented by calculating the polymerization degree of similar patches corresponding to the neighboring LR pixels. The position encoding significantly improved the matching accuracy. Second，a multiscale adaptive fusion module based on channel attention was proposed to improve the feature representation for SR image reconstruction. Experimental results demonstrate that our model outperforms other cutting-edge SR approaches at both 4×and 8×SR. © 2023 Tianjin University. All rights reserved.