Invertible Attention-Guided Adaptive Convolution and Dual-Domain Transformer for Pansharpening

Abstract

Pansharpening is the process of fusing a multispectral (MS) image with a panchromatic image to produce a high-resolution MS (HRMS) image. However, existing techniques face challenges in integrating long-range dependencies to correct locally misaligned features, which results in spatial-spectral distortions. Moreover, these methods tend to be computationally expensive. To address these challenges, we propose a novel detail injection algorithm and develop the invertible attention-guided adaptive convolution and dual-domain Transformer (IACDT) network. In IACDT, we designed an invertible attention mechanism embedded with spectral-spatial attention to efficiently and losslessly extract locally spatial-spectral-aware detail information. In addition, we presented a frequency-spatial dual-domain attention mechanism that combines a frequency-enhanced Transformer and a spatial window Transformer for long-range contextual detail feature correction. This architecture effectively integrates local detail features with long-range dependencies, enabling the model to correct both local misalignments and global inconsistencies. The final HRMS image is obtained through a reconstruction block that consists of residual multireceptive field attention. Extensive experiments demonstrate that IACDT achieves superior fusion performance, computational efficiency, and outstanding results in downstream tasks compared to state-of-the-art methods.