M2Convformer: Multiscale Masked Hybrid Convolution-Transformer Network for Hyperspectral Image Super-Resolution

How to fully utilize spectral correlation to improve spatial resolution is the focus of current hyperspectral image super-resolution (HSI-SR) research. Existing approaches either combine deep models with various advanced attention mechanisms to form an end-to-end framework, or concentrate on the problem of modeling prior estimates of spectral bands and space. While most of these methods are designed for supervised learning with paired labels, they may also benefit from self-supervised learning techniques such as masked autoencoders (MAE). This work focuses on the single hyperspectral image super-resolution problem and develops a multi-scale masked hybrid convolution-transformer framework. The starting point of this work is an attempt to add a random mask to the input signal to reduce the redundancy of the original features, which the model combines with multi-scale representation inference to improve its learning and generalization capabilities. However, we found that simply deploying MAE only for HSI-SR tasks leads to subpar performance. To solve this problem and coordinate with the multi-scale network, we propose a multi-scale inter-channel masking fusion strategy to save computational overhead and bridge the gap between spectral resolution and spatial resolution. Extensive evaluations on three benchmark datasets demonstrate that the proposed method achieves superior performance than state-of-the-art methods.