|
|
Face Group Projects |
Subspace Learning
This project is aimed to understand fundamental problems in pattern analysis and classification, and to develop new learning techniques and algorithms for modeling and classification of real world image, video, audio and speech data. The basic issues are: (1) understanding intrinsically low-dimensional structures or sub-manifolds of patterns of interest embedded in high dimensional data, and (2) discriminating between different patterns. The topics include linear and nonlinear subspace learning, example-based learning, statistical and neural network methods for modeling and classification.
Local Non-Negative Matrix factorization (LNMF). LNMF is a novel method we proposed for learning spatially localized, parts-based subspace representation of visual patterns, in addition to the non-negativity constraint in the standard NMF (Lee and Seung, 1999). It derives set of bases from training images, which not only allows a non-subtractive (part-based) representation of images but also manifests localized features. Experimental results have demonstrated advantages of LNMF over PCA and NMF.
Nonlinear Dimension Reduction. By using a nonlinear support vector regression array and a subsequent nonlinear normalization stage, multi-view face patterns are mapped to a zero mean Gaussian distribution in an invariant feature space. Properties of the nonlinear mappings and the Gaussian face distribution are explored and supported by experiments.
ICA for View Classification . Independent component analysis (ICA) and its variants, are used for the learning task because they are able to take into account higher order statistics required to characterize the view of object. View-subspace representations can be learned in either unsupervised (with view-unlabeled data) or supervised (with view-labeled data) way.
Subspace Modeling for Image Retrieval. This is to extract features for the class of images represented by the positive images provided by subjective RF. PCA is used to reduce both noise contained in the original image features and dimensionality of feature spaces. The method increases the retrieval speed and reduces the memory significantly without sacrificing the retrieval accuracy.
Subspace Modeling of Speaker Variability. Speakers are usually represented by feature vectors in a very high dimension space. PCA and ICA are used to obtain low-dimension representation of speaker variance, including gender and accent.