Nearly-Isoentropic Non-Linear Operators for Interframe Video Coding

• Type: Article in International Conference Proceedings Book
• Date: 1996

Authors

André Puga (Author) FEUP; Artur Pimenta Alves (Author) FEUP

Conference proceedings International

Title: Proceedings Internacional Conference On the Mathematics of Image IMAGETECH96

Internacional Conference On the Mathematics of Image - IMAGETECH96

Atlanta

Other information

Resumo (PT): The objective of this contribution is to introduce an approach to interframe video coding assuming a priori that the source is non Gaussian. To this end, high-order statistics and Generalised Gaussian (GG) modelling are used and a new optimal linear coding problem is identified as a simultaneous diagonalisation of 2nd and 4th order cumulant tensors. This problem, the we named High-Order Karhunen-Loéve Transform (HOKLT) is close to Independent Component Analysis (ICA) previously introduced by P. Comom but it has an additional isoentropic restriction. Using the common available techniques for cumulant tensor diagonalisation, the HOKLT problem can not be, in general, solved exactly. Assuming the impossibility of solving HOKLT problem in the linear group, isoentropic zero-memory non-linearities that can be used to convert between univariate GG distributions were introduced. In particular those non-linearities can be used for the convertion to Gaussian distribution that is an instance of GG. The aggregation of several such non-linearities in a diagonal non-linear operator will be used to convert to a gaussian principal components problem that is solved using classical KLT. Technical and fundamental problems are identified for further refinement of the proposed principle and some possible schemes for the application of this principle are introduced.

Language: Portuguese

Type (Professor's evaluation): Scientific