Higher-order statistics approach to hybrid video coding

• Type: Article in International Conference Proceedings Book
• Date: 1995

Authors

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

Conference proceedings National

Title: Proceedings of the 7th Portuguese Conference on Pattern Recognition

Pages: 841-845

7th Portuguese Conference on Pattern Recognition

Aveiro, 23 - MARÇO - 1995

Other information

Resumo (PT): The "gaussianity" of natural images has historically imposed the conventional second-order statistical approach to the image coding problem. Nowadays, there is a new and nearly stable framework for data analysis, signal processing and, consequently, for source coding. This framework, named higher-order statistics, provides means to analyse non-gaussian and non-linear information sources. In particular it is possible to detect a non-linear structure in signals/sources and it is possible to extend the principal components concept to higher-order objects representing higher-order statistics. In the present paper, we will make the first study, within HOS framework, concerning the two types of images present in the hybrid video coder: natural images (or intraframe) and non-natural images (or interframe). The skewness and kurtosis measures obtained from non-natural images show insignificant deviation from symmetry but deviation from gaussianity in the 4th order measure. So, there are reasons to try new approaches using 4th order information and we will introduce the formulation of the problem of optimal linear transform for symmetric non-gaussian sources. We will name it "Higher-Order Karhunen-Loève Transform" - HOKLT - and it will be seen that it reduces to a tensor diagonalization problem. We introduce the properties of that tensor and, finally, we present some directions for developments based in higher-order statistics.

Language: Portuguese

Type (Professor's evaluation): Scientific

No. of pages: 5