Login

Topics

Found 4 items.
  • Partially Supervised Blind Source Separation
    • Topic: Partially Supervised Blind Source Separation
    • Specification: The goal of Blind Source Separation (BSS) is to separate mixtures of signals (sources) that are observed by several sensors. The term "blind" means that minimum information should be used, and the separation should be based only on general mathematical principles. Independent Component Analysis is a BSS method based on the assumption that the originals signals are statistically independent. BSS methods are widely used as the conditions for being applicable are very general, however, their performance is limited due to lack of information. They also suffer from indeterminacies (e.g. the original order of signals cannot be determined without additive knowledge). BSS methods can be modified in various ways so that some knowledge is used for improving the separation. These methods are often referred to as semi-blind. The goal of this thesis is to develop new semi-blind approaches, especially, for audio source separation, such that are robust to inaccurate information. Also, new methods for side information acquisition (such as through training a deep neural network) can be developed.
    • Type of project: Thesis
    • Duration: 36 months
    • Specialization: FM - Technical Cybernetics
    • Contact person: zbynek.koldovsky@tul.cz
    • Additional information: Scholarship is available,Suitable for students of PhD programme
  • New Models for Independent Component/Vector Extraction
    • Topic: New Models for Independent Component/Vector Extraction
    • Specification: The goal of Blind Source Separation (BSS) is to separate mixtures of signals (sources) that are observed by several sensors. Blind Source Extraction (BSE) aims at extracting only one particular signal. We have been developing methods for Independent Component/Vector Extraction (ICE/IVE) that extract the desired signal blindly based on the assumption that it is statistically independent of the other (background) signals. The ideas of ICE/IVE can be also applied to modify advanced BSS method for BSE. For example, the recently proposed Independent Low-Rank Matrix Analysis (ILRMA) can be modified this way. The goal of this thesis is to develop such modifications and, also, new models such as block-wise determined models or other models considering joint parameters.
    • Type of project: Thesis
    • Duration: 36 months
    • Specialization: FM - Technical Cybernetics
    • Contact person: zbynek.koldovsky@tul.cz
    • Additional information: Scholarship is available,Suitable for students of PhD programme
  • Pushing Fundamental Limitations of Frequency-Domain Blind Source Separation
    • Topic: Pushing Fundamental Limitations of Frequency-Domain Blind Source Separation
    • Specification: The goal of Blind Audio Source Separation (BASS) is to separate signals that are observed by an array of microphones as mixtures of signals originating from individual sources (e.g. speakers). BASS is typically performed in the frequency-domain: The observed signals are divided into frames and each frame is transformed by Discrete Fourier Transform (DFT). This causes some fundamental limitations for the separation accuracy. The signals in the time-domain obey the convolutive mixing, which is, in the frequency domain, approximated as pure multiplication (instantaneous mixing). Circular-convolution effects, framing, the finite length of DFT and so forth impose approximation errors. The goal of this thesis is to develop approaches for minimizing the adverse effects on the accuracy of BASS. Methods of multirate digital signal processing, new models for BSS and/or special sensors can be used to this end.
    • Type of project: Thesis
    • Duration: 36 months
    • Specialization: FM - Technical Cybernetics
    • Contact person: zbynek.koldovsky@tul.cz
    • Additional information: Scholarship is available,Suitable for students of Bachelor programme
  • Blind Source Separation Through Joint Independent Subspace Extraction
    • Topic: Blind Source Separation Through Joint Independent Subspace Extraction
    • Specification: The goal of Blind Source Separation (BSS) is to separate mixtures of signals (sources) that are observed by several sensors. BSS methods based on the assumption that the sources to-be separated are independent have been very popular due to their wide applicability. Blind Source Extraction is a sub-problem of BSS where only one source should be separated from the other background signals. Our group has been developing methods for this problem called Independent Component/Vector Extraction (ICE/IVE). The goal of this thesis is to modify existing ICE/IVE approaches for extracting subspaces of signals (the existing methods extract only one-dimensional signals). The development of novel methods and the computation of Cramer-Rao bound for the problem could be also the goal.
    • Type of project: Thesis
    • Duration: 36 months
    • Specialization: FM - Technical Cybernetics
    • Contact person: zbynek.koldovsky@tul.cz
    • Additional information: Scholarship is available,Suitable for students of PhD programme
General partners
  • ČEZ
    ČEZ
  • Škoda Auto
    Škoda Auto
Partners
  • Preciosa
    Preciosa
  • INISOFT
    INISOFT
  • digades GmbH
    digades GmbH
  • KAUTEX TEXTRON GMBH & CO.
    KAUTEX TEXTRON GMBH & CO.
Schools
  • SPŠ a VOŠ Jičín
    SPŠ a VOŠ Jičín
  • SPŠ Česká Lípa
    SPŠ Česká Lípa
  • SPŠSE a VOŠ Liberec
    SPŠSE a VOŠ Liberec
  • SOŠ, SPŠ Varnsdorf
    SOŠ, SPŠ Varnsdorf
  • SPŠ Mladá Boleslav
    SPŠ Mladá Boleslav