DEEP HYPERSPECTRAL TENSOR COMPLETION JUST USING SMALL DATA

Unlike RGB images available almost everywhere, hyperspectral remote sensing images are not easily obtainable, making big data collection often infeasible for a target task. This would prevent the adoption of the powerful deep learning technology from being applied in solving some challenging problems, e.g., recovering the missing part of a hyperspectral data cube (viewed as a 3-way tensor). This fact induces a series of research works to investigate how to augment the small data, for example, by rotation. We just accept the fact that only small data is available, and propose a radically different view (without augmentation) to address the lacking of big data in hyperspectral remote sensing. Specifically, we show that a deep neural network trained using just small data can still output some useful information to be used in designing regularizer for the ill-posed hyperspectral tensor completion (HTC) problem. Such regularizer is made simple and convex, thereby allowing us to design a fast convex optimization based HTC algorithm, whose superiority is experimentally demonstrated.