Answers are grounded in this paper and cite their sources.
Spectral reconstruction aims to learn the inverse mapping from RGB images to hyperspectral images. However, this remains a challenging ill-posed problem due to the severe information gap and overlapping spectral response ranges among RGB channels. To alleviate the representation-level entanglement caused by this inherent overlap, we propose a new spectral dual matching transformer (SDMT) based on a channel-decoupled representation and cross-channel collaborative fusion framework. Instead of processing the mixed RGB input monolithically, our framework extracts channel-specific features independently to prevent premature spectral crosstalk. Specifically, SDMT incorporates a spectral dual matching self-attention mechanism consisting of a sparse branch and a dense branch. The sparse branch utilizes Top- k filtering to dynamically match highly correlated spectral bands, while the dense branch preserves long-range global correlations. Their features are adaptively fused to enhance key spectral details while maintaining overall background information. Furthermore, a cross-channel attention fusion module (CAFM) is developed to integrate the intermediate hyperspectral cubes from the decoupled channels. By executing cross-attention calculations among channels, CAFM effectively captures complementary information and rich interchannel dependences. Extensive experiments on the NTIRE2020, CAVE and TG1HRSSC benchmarks demonstrate that our proposed method is superior to the existing advanced methods.
No takes yet. Share an insight, caveat, or question.
Sun et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: