Transfer Learning of Mixture Texture Models

0535433 - ÚTIA 2021 RIV CH eng C - Conference Paper (international conference)
Haindl, Michal - Havlíček, Vojtěch
Transfer Learning of Mixture Texture Models.
Computational Collective Intelligence. Cham: Springer Nature Switzerland AG, 2020 - (Nguyen, N.; Hoang, B.; Huynh, C.; Hwang, D.; Trawinski, B.; Vossen, G.), s. 825-837. Lecture Notes in Artificial Intelligence, 12496. ISBN 978-3-030-63006-5. ISSN 0302-9743. E-ISSN 1611-3349.
[International Conference on Computational Collective Intelligence 2020 /12./. Da Nang (VN), 30.11.2020-03.12.2020]
R&D Projects: GA ČR(CZ) GA19-12340S
Institutional support: RVO:67985556
Keywords : Texture modeling * transfer learning * compound random field model * bidirectional texture function
OECD category: Automation and control systems
http://library.utia.cas.cz/separaty/2020/RO/haindl-0535433.pdf

A transfer learning approach for multidimensional parametric mixture random field-based textural representation is introduced. The proposed transfer learning approach allows alleviating the multidimensional mixture models requirement for sufficiently large, but not always available, learning data sets. These compound random field models consist of an underlying structure model that controls transitions between several sub-models, each of them has different characteristics. The structure model proposed is a two-dimensional probabilistic mixture model, either of the Bernoulli or Gaussian mixture type. Local textures are modeled using the fully multispectral three-dimensional Gaussian mixture sub-models. Both presented compound random field models allow the reproduction of, compresses, edits, and enlarges a given measured color, multispectral, or bidirectional texture function (BTF) texture so that ideally, both measured and synthetic textures are visually indiscernible.
Permanent Link: http://hdl.handle.net/11104/0314144