PROJECT TITLE :

A Physics-Based Deep Learning Approach to Shadow Invariant Representations of Hyperspectral Images - 2018

ABSTRACT:

This Project proposes the Relit Spectral AngleStacked Autoencoder, a novel unsupervised feature learning approach for mapping pixel reflectances to illumination invariant encodings. This work extends the Spectral Angle-Stacked Autoencoder therefore that it can learn a shadow-invariant mapping. The strategy is impressed by a deep learning technique, Denoising Autoencoders, with the incorporation of a physics-based mostly model for illumination such that the algorithm learns a shadow invariant mapping while not the necessity for any labelled training knowledge, further sensors, a priori data of the scene or the belief of Planckian illumination. The strategy is evaluated using datasets captured from many totally different cameras, with experiments to demonstrate the illumination invariance of the features and the way they can be used practically to boost the performance of high-level perception algorithms that operate on pictures acquired outdoors.