Search papers, labs, and topics across Lattice.
This paper introduces GSPan, a novel framework for pansharpening that utilizes 2D Gaussian Splatting to represent band-wise residual details as continuous, learnable primitives. By employing a Dual-Stream Hierarchical Interaction architecture with a Spatial-Spectral Interactive Attention module, GSPan effectively integrates low-resolution multispectral and panchromatic images without the need for scale-specific retraining. Experimental results across multiple datasets indicate that GSPan not only achieves state-of-the-art fusion performance but also enhances computational efficiency through its Scale-Decoupled Asymmetric Inference strategy.
GSPan redefines pansharpening by leveraging continuous Gaussian representations, enabling high-quality image fusion across arbitrary scales without retraining.
Pansharpening aims to generate high-resolution multispectral (HRMS) images by fusing low-resolution multispectral (LRMS) and panchromatic (PAN) observations. Most existing deep learning methods treat pansharpening as fixed-grid prediction, which limits scale adaptation. To address this, we propose GSPan, a framework that introduces 2D Gaussian Splatting (GS) into pansharpening. Instead of directly predicting pixels, GSPan represents band-wise residual details as continuous and learnable 2D Gaussian primitives. We design a Dual-Stream Hierarchical Interaction (DSHI) architecture with a Spatial-Spectral Interactive Attention (SSIA) module to estimate these primitives from complementary PAN and MS observations. The predicted primitives are rendered as a residual detail field and injected into the upsampled MS image. This continuous representation allows GSPan to render fused images on arbitrary target sampling grids without scale-specific retraining. It further enables a Scale-Decoupled Asymmetric Inference (SDAI) strategy, which estimates primitives at a reduced resolution and renders the fused image at the target resolution for efficient large-scene pansharpening. Experiments on QuickBird, GaoFen-2, WorldView-3, and WorldView-3-4K datasets show that GSPan delivers state-of-the-art fusion performance. Moreover, SDAI markedly accelerates inference, achieving a favorable trade-off between computational efficiency and fusion quality. Our results demonstrate the potential of continuous Gaussian residual representations as a flexible and scale-decoupled alternative to fixed-grid prediction.
Tsinghua AI