Multiscale implicit neural registration and fusion network for unregistered hyperspectral image fusion

Fusion-based hyperspectral super-resolution methods typically assume strict registration between the two input images. However, in real-world scenarios, spatial misalignment exists between the two modalities, and forced fusion inevitably leads to spectral distortion. To address this, we propose a multi-scale implicit neural registration and fusion network, consisting of encoders, a registration module, a reconstruction module, and a fusion module. The registration module employs global attention and implicit alignment to correct misalignment. The reconstruction module recovers high-frequency details via learnable spectral bases and spatially adaptive refinement. The outputs are adaptively fused through channel attention gating. Through progressive multi-scale refinement, a high-resolution hyperspectral image(HR_HSI) is obtained. Experiments demonstrate that the method stably produces high-quality reconstructions even with unregistered inputs.