Channel Dimension Reduction in Hyper-Spectral Images via Learnable Filters for Vision Tasks

Hyper-spectral images describe data cubes encoding three-dimensional pixel information derived from a continuous frequency spectrum. When the frequency spectrum is discretised and significant gaps between individual frequencies are allowed, the data cube encodes a multi-spectral image. While hyper-spectral and multi-spectral images encode many representative features of objects, events, and signals, many image channels only contain noise. In some applications such as DAS systems, hyper-spectral images can potentially consist of 2500 channels or more with varying signal-to-noise ratios. The research problem in this thesis is stated as channel dimension reduction in hyper-spectral images via learnable filters for vision tasks. The modelling of learnable filters is investigated to select frequency bands or channels for hyper-spectral image construction.