MST-: A Modification of MST++ for Narrow Domain Hyperspectral Reconstruction

Daniil Vladimirov, Daniil Reutsky, Pavel Pischev, Vsevolod Plokhotnyuk, Egor I. Ershov

Abstract: This paper is a technical report of "DiscreteSkishki" team’s solution for ICASSP 2024 Grand Challenge on Hyperspectral Skin Vision. The goal of this competition was to develop human skin spectral reconstruction method using 3 channels in visible (VIS) and a channel in near-infrared ranges (NIR). The feature of this task is a narrow domain of data, which existing methods are not likely to exploit. As a solution we introduce a modification of MST++, by leaving only one SST block instead of three to reduce the number of parameters. Our approach took the second place in the competition.