Joint low-rank dictionary learning for multispectral image spectral super-resolution from training and testing sets

Hyperspectral imaging technology has gained extensive application across diverse fields owing to its rich spectral information. Nonetheless, the high costs associated with data acquisition present a significant barrier to the further advancement of hyperspectral imaging. Recent research indicates that the spectral details of hyperspectral images can be effectively reconstructed from images with lower spectral resolution, such as multispectral images (MSIs), thereby offering a potential path to reducing costs. Despite the remarkable achievements of sparse representation theory across various fields of computer image processing, particularly in spectral super-resolution (SSR), the existing SSR methods based on sparse representation primarily focus on learning a high-spectral-resolution dictionary and a low-spectral-resolution dictionary from the coupled spaces of the training dataset. However, because these dictionaries are exclusively derived from the training dataset, there is an underlying concern that they may not accurately capture the spectral characteristics of the testing data. We introduce a sparse representation-based approach that jointly learns low-rank dictionaries from both the training and testing datasets for MSI spectral super-resolution reconstruction. By constructing common dictionaries that are informed by both the training and testing datasets, our method ensures that these dictionaries are well-suited to represent and effectively model both datasets. Quantitative evaluation and qualitative analysis conducted on simulated and real datasets reveal the superiority of the proposed method over other state-of-the-art methods.