Spectral reconstruction (SR) aims to recover high-quality hyperspectral images (HSIs) from more readily available RGB or multispectral images (MSIs). While supervised SR has shown promising results, it is hindered by the difficulty of collecting abundant, well-registered RGB-HSI or MSI-HSI pairs. Semi-supervised SR (Semi-SR) offers a more practical solution by exploiting plentiful RGBs/MSIs together with limited HSIs. However, existing Semi-SR approaches still suffer from cross-domain discrepancies, cross-modality inconsistency, and unreliable pseudo-labels. To tackle these challenges, we propose a Manifold-aware Teacher-Student Semi-SR (MTSSR) framework, which seamlessly integrates labeled and unlabeled domains through a teacher-student paradigm and memory-efficient consistency learning. At its core, a Flexible Cross-attention Spectral Reconstruction (FCSR) network extracts scene-related spatial cues via customized self-attention and models scene-agnostic priors through dynamic quantization, thereby enhancing spectral fidelity. Furthermore, a manifold-aware dimensionality analysis derives a latent space that jointly captures spatial and spectral structures across modalities. This enables a manifold-aware alignment loss to enforce cross-modality consistency and a manifold-aware contrastive loss to progressively refine pseudo-label reliability. In addition, we develop a Threshold-adjusted Memory Bank Update (TMBU) strategy, which generates reliable negative samples by storing network-driven representations instead of memory-consuming HSIs, significantly reducing memory consumption. Extensive experiments on three visual and two remote sensing benchmarks demonstrate that MTSSR consistently outperforms state-of-the-art SR methods, achieving robust and memory-efficient spectral reconstruction.

Existing supervised spectral reconstruction (SR) methods adopt paired RGB images and hyperspectral images (HSIs) to drive the overall paradigms. Nonetheless, in practice, "paired" requires higher device requirements such as specific well-calibrated dual cameras or more complex and exact registration processes among images with different time phases, widths, and spatial resolution. To tackle the above challenges, we propose a flexible uncertainty-aware unsupervised SR paradigm, which dynamically establishes the forceful and potent constraints with RGBs for driving unsupervised learning. As a specific plug-and-play tail in our paradigm, the uncertainty-aware saliency alignment module (USAM) calculates pixel- and spectralwise information entropy for uncertainty estimation, which attempts to represent the corresponding reflectivity or radiance to the light among different objects in various scenes, forcing the paradigm to adaptively explore the scene-agnostic prominent features. Furthermore, a progressively parallel network under our unsupervised paradigm is conducted to excavate discriminate structural and semantic priors of RGBs to assist in recovering dependable HSIs: 1) a learnable rank-guided structural representation (LRSR) flow is leveraged to characterize the latent structural priors via excavating nonzero elements in the full-rank matrix and further preserve evident boundaries in HSIs; and 2) a coarse-to-fine bandwise semantic perception (CBSP) flow is conducted to propagate perceptual bandwise affinity for aggregating and strengthening intrinsic interband dependencies, and further extract delicate semantic priors, which can recover plentiful contiguous spectral information in HSIs. Comprehensive quantitative and qualitative experimental results on three visual and two remote sensing benchmarks have shown the superiority and robustness of our method. We also conducted nine existing SR methods in our unsupervised paradigm to recover HSIs without any manual intervention, which proves the generality of our paradigm to some extent.

Currently, leading methods for spectral super-resolution (SSR) depend heavily on constructing diverse network architectures in a heuristic manner, in order to learn a full mapping from the RGB image to its corresponding hyperspectral image (HSI). Despite promising results in reconstruction performance, significant challenges remain with respect to model interpretation and the capture of long-range dependencies. In response to these issues, based on a comprehensive exploration of the physical imaging mechanism between spectral response curve (SRC) and HSI, we have developed a novel model-driven degradation-aware unfolding network (DAUNet) in an iterative way. Besides, the learning process is explicitly integrated with the intrinsic generation mechanism of the SSR task. To be specific, we unfold each step into a degradation-aware gradient decent (DAGD) module and a proximal mapping module (PMM), using the framework of maximum a posteriori (MAP) theory. Additionally, to introduce more discriminative learning capabilities to our network, we have further enhanced the PMM architecture by incorporating a fine-grained multihead spectral-wise transformer (FMST) block, which improves global feature representation compared to the channel-wise transformer block. Extensive experiments over several spectral datasets finely demonstrate the superior performance of our method beyond the current representative state-of-the-art (SOTA) SSR methods.

A significant challenge of spectral reconstruction (SR) task is the lower performance reconstructed in foreground regions compared to background regions, which can be attributed to the marked difference in diversity of objects and disparity of adjacent scene characteristics. Moreover, the reconstruction of edge regions is often fraught with substantial errors due to the transitional nature of these regions, an issue conventional single convolutional neural networks (CNNs) and transformers struggle to handle. To address these challenges, we introduce the residual mask in cascaded convolutional transformer (RC2T) to iteratively improve the reconstruction of hyperspectral images (HSIs). Specifically, we propose a residual-predict mask generator (RMG) to generate a residual mask that retains band properties to separate feature with different complexities. Meanwhile, to achieve band expansion of mask features within the autoencoder, we approximate it to a Markov process and exploit the multistage spectral-aware Markov transfer (MMT) for its lightweight implementation. Next, we introduce the parallel convolutional multihead self-attention module (PSM), in which CNN runs parallel to the transformer to handle simple and complex features separately. Additionally, the residual mask loss function uses the established relationship between complexity of feature and reconstruction accuracy to generate residual mask in a self-supervised manner for providing complex high-frequency prior. We have validated our approach using three published datasets (NTIRE 2020 “Clean” track, NTIRE 2022, and CAVE). Additionally, we also conducted experiments with the proposed method on remote sensing dataset grss_dfc_2018 and a satellite-borne remote sensing dataset, achieving optimal performance. The experimental results demonstrate that our RC2T method is state-of-the-art (SOTA) in the field of SR.

Spectral reconstruction (SR) aims to recover the hyperspectral images (HSIs) from the corresponding RGB images directly. Most SR studies based on supervised learning require massive data annotations to achieve superior reconstruction performance, which are limited by complicated imaging techniques and laborious annotation calibration in practice. Thus, unsupervised strategies attract attention of the community, however, existing unsupervised SR works still face a fatal bottleneck from low accuracy. Besides, traditional CNN-based models are good at capturing local features but experience difficulty in global features. To ameliorate these drawbacks, we propose an unsupervised SR architecture with strong constraints, especially constructing a novel Masked Transformer (MFormer) to excavate latent hyperspectral characteristics to restore realistic HSIs further. Concretely, a Dual Spectral-wise Multi-head Self-attention (DSSA) mechanism embedded in transformer is proposed to firmly associate multi-head and channel dimensions and then capture the spectral representation in the implicit solution spaces. Furthermore, a plug-and-play Mask-guided Band Augment (MBA) module is presented to extract and further enhance the band-wise correlation and continuity to boost the robustness of the model. Innovatively, a customized loss based on the intrinsic mapping from HSIs to RGB images and the inherent spectral structural similarity is designed to restrain spectral distortion. Extensive experimental results on three benchmarks verify that our MFormer achieves superior performance over other state-of-the-art supervised and unsupervised methods under a no-label training process equally.