Fundamentals of blind source separation

• Grant number: 13/07391-5
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: August 01, 2013
• End date: July 31, 2015
• Field of knowledge: Engineering - Electrical Engineering - Telecommunications
• Principal Investigator: Romis Ribeiro de Faissol Attux
• Grantee: David Felice Falivene Baptista
• Host Institution: Faculdade de Engenharia Elétrica e de Computação (FEEC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Associated scholarship(s): 14/25335-8 - Nonlinear blind source separation and chemical sensors, BE.EP.IC

Abstract

The problem of blind source separation (BSS) corresponds, in simple terms, to the notion of extracting information signals from measurements in which they are mixed and/or distorted. This problem arises in fields as diverse as those of satellite image processing and auditory scene analysis. In formal terms, in order the BSS problem be solved, it is necessary to define a separating structure and a criterion to optimize its free parameters. Classically, the separating system is assumed to be linear and instantaneous and the sources are considered as being mutually independent, which establishes a connection between BSS and ICA (independent component analysis). There are several criteria that can be employed to perform BSS via ICA. Among them, we highlight those that make use of the concept of nongaussianity (e.g. based on kurtosis and negentropy), those that employ estimates of the mutual information between outputs and those founded on the maximum-likelihood approach. In this work, the main methodologies for performing BSS using ICA will be studied and, afterwards, a comparative analysis will be carried out within several simulation setups. The main goal of this effort is to form a young researcher in this important research field and to bring together analytical elements potentially useful to those interested in using BSS / ICA in practical contexts. (AU)