Deep low-rank tensor embedded network for hyperspectral image super-resolution

Recent efforts have witnessed significant progress in deep-learning-based hyperspectral image super-resolution (HSISR). However, most existing methods focus solely on spatial or spectral exploration, while lacks enough consideration of the intrinsic correlation between these aspects. This oversight limits the potential for collaborative optimization, leading to suboptimal feature representations of HSI. Moreover, they mainly engaged in super-resolve the pixel-wise spatial details, neglecting the vital spectral consistency. To mitigate these issues, this paper proposed LRTENet, a novel deep low-rank tensor embedding network for HSISR, which effectively bridges the optimization gap between spatial and spectral features with well-defined low-rank tensor decomposition. Specially, we introduce a low-rank embedding module (LREM) to extract low-rank dependencies across multiple directions facilitating a holistic mapping by adaptively integrating these tensors. This enables our model to generate discriminative spatial-spectral representations for accurate reconstruction. Furthermore, to better preserve the spectral consistency, we incorporate LREM after upsample operation to progressively refine and correct spectral distortion. Extensive experiments demonstrate that LRTENet achieves superior spatial reconstruction and spectral preservation performance, outperforming state-of-the-art methods on various benchmarks, including Chikusei, CAVE, and Pavia.