Segmentierung von Gesichtsbildern

Martin Hirzer

Publikation: StudienabschlussarbeitMasterarbeitForschung

Abstract

Seit der Einführung persönlicher Dokumente in elektronischer Form ist die Analyse von Passbildern zu einem wichtigen Forschungsbereich geworden. Um für elektronische Dokumente geeignet zu sein, müssen Passbilder eine Reihe von Qualitätsanforderungen, die von der Internationalen Zivilluftfahrt Organisation (ICAO) definiert sind, erfüllen. Einige dieser Kriterien betreffen das ganze Bild, andere wiederum beziehen sich nur auf bestimmte Regionen im Bild. Zur Kontrolle solch regionenspezifischer Bedingungen ist eine Lokalisierung der entsprechenden Bildbereiche im Vorhinein notwendig. In dieser Arbeit präsentieren wir eine Methode zur automatischen Segmentierung von Gesichtsbildern in Farbe. Der entwickelte Algorithmus soll anschließend in einem automatischen Validierungsgerät für Passbilder eingesetzt werden. Zum Erreichen dieses Ziels verwenden wir einen wissensgesteuerten Ansatz. Dieser basiert auf einem Total-Variation-Modell das es uns erlaubt, Vorwissen über den Inhalt typischer Passbilder in den Segmentierungsprozess einzubinden. Da eine der Qualitätsanforderungen für Passbilder ein gleichmäßiger Bildhintergrund ist, entwickelten wir zusätzlich einen Hintergrundklassifikator, der, ausgehend vom Segmentierungsergebnis, den Bildhintergrund als homogen oder nicht homogen kennzeichnet. Dadurch können Bilder mit nicht homogenem Hintergrund bei der Überprüfung schon zu einem frühen Zeitpunkt zurückgewiesen werden, und müssen nicht den gesamten Validierungsprozess durchlaufen. Schließlich haben wir auch zahlreiche Experimente auf zwei verschiedenen Datensätzen von Gesichtsbildern durchgeführt, um unseren Algorithmus zu evaluieren. Die erzielten Resultate zeigen, dass unsere Methode sehr robust ist und andere Ansätze zur Segmentierung von Gesichtsbildern, nämlich ein Expertensystem und einen AdaBoost-Klassifikator, übertrifft.
Titel in ÜbersetzungSegmentierung von Gesichtsbildern
Originalspracheenglisch
PublikationsstatusVeröffentlicht - 2008

Fingerprint

Civil aviation
Classifiers
Inspection
Adaptive boosting
Image segmentation
Expert systems
Color
Experiments

Fields of Expertise

  • Information, Communication & Computing

Dies zitieren

Hirzer, M. (2008). Segmentation of Face Images.

Segmentation of Face Images. / Hirzer, Martin.

2008. 101 S.

Publikation: StudienabschlussarbeitMasterarbeitForschung

Hirzer, M 2008, 'Segmentation of Face Images'.
Hirzer M. Segmentation of Face Images. 2008. 101 S.
Hirzer, Martin. / Segmentation of Face Images. 2008. 101 S.
@phdthesis{125ac84e00e44ce68f71a13737021273,
title = "Segmentation of Face Images",
abstract = "Since the introduction of electronic personal documents in recent years the analysis of passport photographs has become an important field of research. Such photographs have to fulfill a set of minimal quality requirements defined by the International Civil Aviation Organization (ICAO). As some of the specified requirements are related to certain image regions only, these regions must be located in advance. In this work an unsupervised segmentation method for color face images is presented. The developed tool is intended to be part of an automatic passport photograph inspection framework. Our focus is on knowledge based image segmentation. Hence we developed a total variation model that allows us to incorporate prior knowledge about typical passport photographs into the segmentation process. Since uniformity of the image background is one quality requirement defined by ICAO, our tool also contains a background classifier, which decides whether the background region is uniform or not. This enables the inspection framework to reject photographs with a non-uniform background at an early stage. We have conducted several experiments on face images from two different datasets in order to evaluate the performance of our algorithm. The obtained results demonstrate that our method is fairly robust and outperforms other methods targeted at the same problem, in particular an expert system and an AdaBoost classifier.",
author = "Martin Hirzer",
year = "2008",
language = "English",

}

TY - THES

T1 - Segmentation of Face Images

AU - Hirzer, Martin

PY - 2008

Y1 - 2008

N2 - Since the introduction of electronic personal documents in recent years the analysis of passport photographs has become an important field of research. Such photographs have to fulfill a set of minimal quality requirements defined by the International Civil Aviation Organization (ICAO). As some of the specified requirements are related to certain image regions only, these regions must be located in advance. In this work an unsupervised segmentation method for color face images is presented. The developed tool is intended to be part of an automatic passport photograph inspection framework. Our focus is on knowledge based image segmentation. Hence we developed a total variation model that allows us to incorporate prior knowledge about typical passport photographs into the segmentation process. Since uniformity of the image background is one quality requirement defined by ICAO, our tool also contains a background classifier, which decides whether the background region is uniform or not. This enables the inspection framework to reject photographs with a non-uniform background at an early stage. We have conducted several experiments on face images from two different datasets in order to evaluate the performance of our algorithm. The obtained results demonstrate that our method is fairly robust and outperforms other methods targeted at the same problem, in particular an expert system and an AdaBoost classifier.

AB - Since the introduction of electronic personal documents in recent years the analysis of passport photographs has become an important field of research. Such photographs have to fulfill a set of minimal quality requirements defined by the International Civil Aviation Organization (ICAO). As some of the specified requirements are related to certain image regions only, these regions must be located in advance. In this work an unsupervised segmentation method for color face images is presented. The developed tool is intended to be part of an automatic passport photograph inspection framework. Our focus is on knowledge based image segmentation. Hence we developed a total variation model that allows us to incorporate prior knowledge about typical passport photographs into the segmentation process. Since uniformity of the image background is one quality requirement defined by ICAO, our tool also contains a background classifier, which decides whether the background region is uniform or not. This enables the inspection framework to reject photographs with a non-uniform background at an early stage. We have conducted several experiments on face images from two different datasets in order to evaluate the performance of our algorithm. The obtained results demonstrate that our method is fairly robust and outperforms other methods targeted at the same problem, in particular an expert system and an AdaBoost classifier.

M3 - Master's Thesis

ER -