We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

eess.IV

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Electrical Engineering and Systems Science > Image and Video Processing

Title: Improving Autoencoder Training Performance for Hyperspectral Unmixing with Network Reinitialisation

Abstract: Neural networks, in particular autoencoders, are one of the most promising solutions for unmixing hyperspectral data, i.e. reconstructing the spectra of observed substances (endmembers) and their relative mixing fractions (abundances), which is needed for effective hyperspectral analysis and classification. However, as we show in this paper, the training of autoencoders for unmixing is highly dependent on weights initialisation; some sets of weights lead to degenerate or low-performance solutions, introducing negative bias in the expected performance. In this work, we experimentally investigate autoencoders stability as well as network reinitialisation methods based on coefficients of neurons' dead activations. We demonstrate that the proposed techniques have a positive effect on autoencoder training in terms of reconstruction, abundances and endmembers errors.
Subjects: Image and Video Processing (eess.IV); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
MSC classes: 68T07
ACM classes: I.4
Cite as: arXiv:2109.13748 [eess.IV]
  (or arXiv:2109.13748v3 [eess.IV] for this version)

Submission history

From: Michał Romaszewski [view email]
[v1] Tue, 28 Sep 2021 14:07:24 GMT (1727kb,D)
[v2] Mon, 21 Mar 2022 08:17:26 GMT (2630kb,D)
[v3] Tue, 12 Apr 2022 08:36:53 GMT (2628kb,D)

Link back to: arXiv, form interface, contact.