Earth Mover's Distance based Local Discriminant Basis (with B. Marchand), in Multiscale Signal Analysis and Modeling (X. Shen and A. I. Zayed, eds.), Lecture Notes in Electrical Engineering, Chap. 12, pp. 275-294, Springer, 2013.
Abstract
Local discriminant Basis (LDB) is a tool to extract useful features for
signal and image classification problems. Original LDB methods rely on the
time-frequency energy distribution of classes or empirical
probability densities, with some information theoretic measure (such as
Kullback-Leibler divergence) for feature selection. Depending on the
problem, energy distributions may not provide the best information for
classification. Further, training set sizes and accuracy in the computed
empirical probability density functions (epdfs) may hinder the learning process
.
To improve these deficiencies and provide a more data adaptive algorithm,
we propose the use of signatures and Earth Mover's Distance (EMD). Signatures
and EMD provide a data adaptive statistic that is more descriptive than the
distribution of energies and more robust than an epdf-based approach.
In this article, we first review LDB and EMD, and then outline how they can
be incorporated into a fast EMD based LDB algorithm. We then demonstrate
the capabilities of our new algorithm in comparison to both energy distribution
and epdf-based LDB algorithms using four different classification problems
using synthetic datasets.
Get the full paper: PDF file.
Get the official version via doi:10.1017/978-1-4614-4145-8_12.
Please email
me if you have any comments or questions!
Go
back to Naoki's Publication Page