Multiscale segmentation method for small inclusion detection in 3D industrial computed tomography

Gerald Zauner, Bernhard Harrer, Daniel Angermaier, Johann Kastner

Publikation: Beitrag in Buch/Bericht/TagungsbandKonferenzbeitragBegutachtung

Abstract

In this paper a new segmentation method for highly precise inclusion detection in 3D X-ray computed tomography (CT), based on multiresolution denoising methods, is presented. The aim of this work is the automatic 3D-segmentation of small graphite inclusions in cast iron samples. Industrial X-ray computed tomography of metallic samples often suffers from imaging artifacts (e.g. cupping effects) which result in unwanted background image structures, making automated segmentation a difficult task. Additionally, small spatial structures (inclusions and voids) are generally difficult to detect e.g. by standard region based methods like watershed segmentation. Finally, image noise (assuming a Poisson noise characteristic) and the large amount of 3D data have to be considered to obtain good results. The approach presented is based on image subtraction of two different representations of the image under consideration. The first image represents the low spatial frequency content derived by means of wavelet filtering based on the 'á trous' algorithm (i.e. the 'background' content) assuming standard Gaussian noise. The second image is derived by applying a multiresolution denoising scheme based on 'platelet'-filtering, which can produce highly accurate intensity and density estimates assuming Poisson noise. It is shown that the resulting arithmetic difference between these two images can give highly accurate segmentation results with respect to finding small spatial structures in heavily cluttered background structures. Experimental results of industrial CT measurements are presented showing the practicability and reliability of this approach for the proposed task.

OriginalspracheEnglisch
TitelOptical Measurement Systems for Industrial Inspection V
AuflagePART 2
DOIs
PublikationsstatusVeröffentlicht - 2007
VeranstaltungOptical Measurement Systems for Industrial Inspection V - Munich, Deutschland
Dauer: 18 Jun 200722 Jun 2007

Publikationsreihe

NameProceedings of SPIE - The International Society for Optical Engineering
NummerPART 2
Band6616
ISSN (Print)0277-786X

Konferenz

KonferenzOptical Measurement Systems for Industrial Inspection V
Land/GebietDeutschland
OrtMunich
Zeitraum18.06.200722.06.2007

Fingerprint

Untersuchen Sie die Forschungsthemen von „Multiscale segmentation method for small inclusion detection in 3D industrial computed tomography“. Zusammen bilden sie einen einzigartigen Fingerprint.

Zitieren