Frequency Domain Matters: An Unsupervised Residual Information Injection Network for Pansharpening
Yang Liu, Wei Wang, Weihe Li · 2025
Pansharpening predicts high-resolution multispectral (HRMS) images by fusing high-resolution panchromatic (PAN) images and multispectral low-resolution (LRMS) images. Spatial-domain pansharpening methods often generate HRMS images with insufficient high-frequency details. Some studies have endeavored to improve pansharpening performance in the frequency domain. Due to the lack of high-resolution multispectral (HRMS) images for supervised training, current frequency domain methods often resort to downsampling full-resolution LRMS-PAN images while using full-resolution LRMS images as ground truth. This approach enables the network to be trained in a fully supervised manner at a reduced resolution. However, models trained on these downsized images tend to underperform when applied to full-resolution target images. This paper presents URIIN, an unsupervised residual information injection network, designed for pansharpening by applying appropriate loss function constraints to full-resolution PAN and LRMS images to generate HRMS images. We present the Frequency Injection Module (FIM), engineered to extract high-frequency residuals from PAN images in the frequency domain and inject them into MS images. Additionally, we utilize a one-shot guided online learning strategy to expedite network training while ensuring pansharpening performance. Comprehensive experiments carried out on multiple datasets substantiate that our method is competitive with the existing state-of-the-art methods in this field.