Mosaic-Aware Channel-Enhanced Global Spectral Modeling for Hyperspectral Reconstruction from Mosaic Images

This paper reports our solution to Track 1 of the Hyper-Object Challenge at the ICASSP 2026 Signal Processing Grand Challenges, which aims to reconstruct high-fidelity hyperspectral images from single-channel mosaic inputs. The main challenge lies in recovering dense spectral information from sparsely and interleaved spectral observations. To address this challenge, we propose a mosaic-aware channel-enhanced global spectral modeling method, i.e., MC-GSM, which integrates shallow mosaic adaptation, channel-wise spectral refinement, and global spectral dependency modeling. Experimental results on the official benchmark demonstrate that our solution achieves competitive reconstruction performance and ranks 2nd on the official leaderboard.