Dynamic deep subspace residual regularizer for hyperspectral image super-resolution

Due to the limitations of imaging mechanisms, hyperspectral image super-resolution technology that fuses high-spatial resolution multispectral images has attracted much attention in recent years. Deep Image Prior is widely used in hyperspectral image super-resolution because of its unsupervised advantage. However, the static noise input of Deep Image Prior is deficient in structural information, which impedes the effective steering of the super-resolution process. To solve this problem, we propose a dynamic deep subspace framework. First, the original data is projected into the subspace domain through spectral low-rank subspace decomposition to utilize the spectral low-rank information while preserving the spatial structure of data, then the subspace representation coefficients are dynamically updated and regarded as input of Deep Image Prior for guiding the super-resolution process. Besides, to fully exploit the intrinsic structural prior in hyperspectral data, we design a dynamic deep subspace residual regularizer in the proposed framework. Specifically, a residual structure is designed using the deep subspace network. The obtained subspace residual coefficients exhibit stronger smoothness and low-rank properties than the original data and subspace coefficients, and can flexibly combine hand-crafted priors such as total variation and low-rank constraints. Experiments on four simulated datasets and one real dataset show that our method achieves competitive performance compared with existing state-of-the-art methods both qualitatively and quantitatively.