Cross-Constraint Regularization Unfolding Network for Pan-Sharpening

Deep unfolding networks (DUNs) have revitalized the field of pan-sharpening with their exceptional performance and interpretability. However, existing DUNs suffer from inadequate integration of priors and insufficient feature alignment, which renders them highly sensitive to feature variations in low-resolution multispectral (LRMS) and panchromatic (PAN) images and ultimately constrains the improvement of reconstruction performance. To address this issue, we propose a cross-constraint regularization unfolding network (CRUN) for pan-sharpening, which decomposes the fusion process into two complementary optimization tasks: a spectral super-resolution branch and a spatial super-resolution branch. Each branch focuses on the feature representation of its respective modality, with tasks providing additional a prior regularization to the other. The proposed CRUN incorporates a customized multiscale degradation perception module (MSDM) to enhance data fidelity and spatial-spectral consistency. Moreover, a deep cross-constraint regularization module (DCRM) with offset-guided cross-attention is introduced to reduce inter-modality discrepancies and spatial misalignments. As a result, the proposed CRUN can capture modality-specific features, ensure spatial-spectral alignment, enhance cross-modal interaction, while maintaining interpretability and yielding accurate and reliable pan-sharpening results. Qualitative assessments confirm that CRUN produces images with rich textual details and minimal spectral distortion. It shows that CRUN consistently outperforms the state-of-the-art methods. In reduced-resolution scene across GaoFen-2, WorldView-II, and WorldView-III datasets, CRUN achieves an average reduction of 16.69% in ERGAS, and improvements of 0.68% in SSIM, 3.33% in SCC, and 6.69% in PSNR. In full-resolution assessment on GaoFen-2 dataset, CRUN attains a QNR value of 0.8553, corresponding to a 7.26% enhancement compared to the average value.