Location

Williamsburg, VA

Start Date

1-1-1986 12:00 AM

Description

If the data produced by a digital imaging system is univariate, i.e. if just one scalar measurement (temperature, density, etc.) is made at each pixel, then standard image processing methods for extracting information from this univariate data are well known and relatively simple to apply [1]. Indeed, most of these methods ultimately reduce to information extraction via a visual inspection of an enhanced or restored image of the data displayed in shades of gray or in pseudocolor. What if the imaging system produces multivariate data i.e. what if an array (or vector) of data is measured at each pixel? In this case some information can be extracted with a visual inspection of each image component. However, it is intuitive that such an approach is fundamentally limited because it is univariate and does not account for the inherently high component-to-component correlation typically found in multivariate images. Instead, what is needed is a more comprehensive approach in which the multivariate data is processed in a space whose dimension matches that of the data. Information extraction via a visual inspection of the data can then take place after some arithmetic processing and statistical decisions have been applied to estimate and remove the multivariate correlation and thereby effectively reduce the dimensionality of the data without significantly reducing its information content.

Book Title

Review of Progress in Quantitative Nondestructive Evaluation

Volume

5A

Chapter

Chapter 2: Inversion, Imaging and Reconstruction

Section

Imaging and Reconstruction

Pages

417-428

DOI

10.1007/978-1-4615-7763-8_43

Language

en

File Format

application/pdf

Share

COinS
 
Jan 1st, 12:00 AM

Information Extraction from Multivariate Images

Williamsburg, VA

If the data produced by a digital imaging system is univariate, i.e. if just one scalar measurement (temperature, density, etc.) is made at each pixel, then standard image processing methods for extracting information from this univariate data are well known and relatively simple to apply [1]. Indeed, most of these methods ultimately reduce to information extraction via a visual inspection of an enhanced or restored image of the data displayed in shades of gray or in pseudocolor. What if the imaging system produces multivariate data i.e. what if an array (or vector) of data is measured at each pixel? In this case some information can be extracted with a visual inspection of each image component. However, it is intuitive that such an approach is fundamentally limited because it is univariate and does not account for the inherently high component-to-component correlation typically found in multivariate images. Instead, what is needed is a more comprehensive approach in which the multivariate data is processed in a space whose dimension matches that of the data. Information extraction via a visual inspection of the data can then take place after some arithmetic processing and statistical decisions have been applied to estimate and remove the multivariate correlation and thereby effectively reduce the dimensionality of the data without significantly reducing its information content.