Loading...
| Friend's email | |
| Your name | |
| Your email | |
| enter code | |
This page was sent successfuly
- Type of Document: Ph.D. Dissertation
- Language: Farsi
- Document No: 50926 (05)
- University: Sharif University of Technology
- Department: Electrical Engineering
- Advisor(s): Babaie-Zadeh, Massoud
- Abstract:
- Blind Source Separation (BSS) is a challenging task in signal processing which aims to separate sources from their mixtures when no information is available about the sources or the mixing system. Different approaches have already been proposed for source separation.However, during the last decade, new approaches based on multimodal nature of phenomena have been proposed for source separation. Different aspects of a multimodal phenomenon can be measured by means of different instruments where each of the measured signals is called a modality of that phenomenon. Although the modalities are different signals with different features, due to the same physical origin, they usually have some similarities and correlations.In this thesis, we have used the similarities among the modalities for separating sources from different mixtures. Four new data fusion approaches have been proposed. The proposed approaches are based on the similarity of the shared factors, where the shred factors result from Nonnegative Matrix Factorization (NMF) of the power spectrogram matrices of the modalities. The proposed approaches are: multimodal soft coupling, multimodal preclustering,multimodal post-clustering and probabilistic multimodal factorization. These four approaches are then used for source separation. For applying these four approaches for source separation, proper multichannel models are needed for modeling the mixtures.The multimodal soft coupling approach is applied to convolutive source separation. In this way, we choose the multichannel model of the Soft NMCF algorithm as the basic multichannel model. The penalty term of the algorithm is replaced by the proposed penalty term. In this way, the shared factors of the mixtures are coupled in a multimodal soft manner. In this new approach, the problem of tending the shared factors to zero, which exists in the Soft NMCF approach, no more occurs.We have also applied the multimodal clustering approaches (pre-clustering and postclustering approaches) for single channel source separation.Clustering of the basis vectors is an important step in single microphone source separation, and the proposed multimodal clustering approaches have the ability of the clustering of the basis vectors. This is confirmed by the simulation results. We have also used multimodal pre-clustering for separating timevarying mixtures. For this propose, in the first step, we have derived a proper multichannel model for time-varying mixtures. In this new model, a four-layer NMF model was used for the sources with time-varying coefficients. The first layer of the modal is a circulant matrix and contains information about the time-varying coefficients. In the second step,the proposed multimodal pre-clustering was applied in the model and new algorithm is derived for separating time-varying mixtures. The simulation results confirm the quality of the proposed algorithms.The problem of the probabilistic factorization of the modalities in the Bayesian ramework is also studied in this thesis. In this problem, the data matrix of one modality is factorized using the information extracted from the other modality. In this approach the shared factors are coupled with each other using their Probability Density Function (PDF) where the variance of the PDF is unknown. So the variance of the model is also estimated along with other parameters. The simulation results confirm the effectiveness of the proposed approaches
- Keywords:
- Blind Sources Separation (BSS) ; Non-Negative Matrix Factorization (NMF) ; Multi-Modal Data ; Shared Factors ; Bayesian Framework
Digital Object List
-
محتواي کتاب
- view
Bookmark