Deep Diffusion Processes for Active Learning of Hyperspectral Images

TL;DR

This paper introduces a novel active learning approach for hyperspectral images that integrates deep variational autoencoders with graph diffusion processes, achieving improved labeling efficiency and performance.

Contribution

It combines deep learning with diffusion geometry for active learning in hyperspectral imaging, a novel integration not previously explored.

Findings

Strong performance on real hyperspectral datasets

Effective feature extraction via deep variational autoencoders

Improved labeling efficiency through diffusion-based queries

Abstract

A method for active learning of hyperspectral images (HSI) is proposed, which combines deep learning with diffusion processes on graphs. A deep variational autoencoder extracts smoothed, denoised features from a high-dimensional HSI, which are then used to make labeling queries based on graph diffusion processes. The proposed method combines the robust representations of deep learning with the mathematical tractability of diffusion geometry, and leads to strong performance on real HSI.