| Fusion of two tightly coupled signals from different sources is addressed in this thesis. In the Bayesian inference framework, fusion of two signals is formulated as estimation of their joint probability. Usually this joint probability of tightly coupled signals from different sources is difficult to estimate directly. To overcome this difficulty, a new approach, based on the maximum entropy principle, is introduced. A “simplicity” criterion and a maximum mutual information criterion are also developed to guide the selection of two dimension-reduction mapping functions used in the proposed approach. We have shown that in the case of linear transformations of Gaussian random vectors, the proposed method is equivalent to the linear canonical correlation analysis method.; Based on the proposed method, two types of fused-hidden Markov models (fused-HMMs) are developed for fusion of two tightly coupled time series. In both types of fused-HMMs, the time series is modeled by an HMM individually, and the two HMMs are fused together with the proposed fusion modal. In a simulation and a bimodal speaker verification experiment, the proposed method significantly reduces the recognition errors.; Finally, kernel canonical correlation analysis (CCA) is developed to model nonlinear or high-order correlations between signals from two sources. Kernel CCA uses kernel principal component analysis (PCA), which elegantly combines a nonlinear transformation and linear PCA into a one-step calculation, so as to avoid the computational burden of high/infinite-dimensional nonlinear transformations. |
