An Expectation-Maximization Algorithm With Local Adaptivity for Image Analysis
Abstract:
In this talk, we present an Expectation-Maximization (EM) algorithm with local adaptivity that can combine statistical inference and geometric information. In particular we apply our algorithm to image segmentation and classification. The idea is to couple global statistics extracted from an efficient statistic model, such as the Gaussian mixture model (GMM), with local statistical and geometrical information, such as local orientation and anisotropy. The combined information is used to design an optimal adaptive local filtering strategy that can both improve robustness and preserve fine features. A multi-resolution, iterative strategy can also be employed to improve both local and global statistical estimates. This framework can be extended easily by taking into account other inferred information/quantities and statistical methods/models.
Applied Mathematics Seminar