Meta-Learning with Dual-Branch Fusion Transformer for Hyperspectral Image Super-Resolution

Unsupervised learning-based methods have gained growing momentum in the field of hyperspectral image (HSI) super-resolution (SR) as they do not rely on the supervision of simulated input-output pairs. However, such methods not only require high-performance computational resources but are also prone to overfit. To address this issue, this paper proposes a meta-learning-based HSI-SR method (MetaHSR) that learns the knowledge of a specific HSI via only one model update iteration.During training, knowledge distillation is utilized as a regularization term to avoid overfitting.Furthermore, a Dual-Branch Fusion Transformer (DBFT) is designed, where local, global, and cross-modal dependencies are learned with convolutions, self-attention, and cross-modal attention, respectively. Extensive experiments demonstrate the superiority of the proposed MetaHSR with respect to state-of-the-art methods.