HSCT: HIERARCHICAL SELF-CALIBRATION TRANSFORMER FOR HYPERSPECTRAL IMAGE SUPER-RESOLUTION

Hyperspectral image (HSI) super-resolution is to improve the spatial resolution while preserving spectral fidelity. Existing CNN- and Transformer-based methods face challenges in simultaneously capturing multi-scale local and global features and maintaining spectral accuracy. To address these issues, in this paper, the Hierarchical Self-Calibration Transformer (HSCT) is proposed for HSI super-resolution, combining the merits of CNNs and Transformers in a multi-stage framework. Specifically, CNNs are utilized for local feature extraction, leveraging inductive biases to enrich feature representations, while Transformers focus on global feature extraction to model complex and global dependencies. A variable Window-based Self-Attention with window shifting is designed to extract multi-scale spatial features, while a Channel Self-Attention refines spectral features to ensure fidelity, parallel integration of which enables efficient spatial-spectral feature learning. Additionally, Self-Calibration Convolution and Residual Connections are integrated to improve feature representations and model stability. Extensive experiments demonstrate the outperformance of the proposed HSCT over representative traditional and state-of-the-art deep learning-based methods, both visually and quantitatively.