TU Darmstadt / ULB / TUbiblio

Estimating Motion from a Single Blurry Image

Gast, Jochen (2014)
Estimating Motion from a Single Blurry Image.
Technische Universität Darmstadt
Masterarbeit, Bibliographie

Kurzbeschreibung (Abstract)

Estimating motion from a single image source is a heavily ill-posed problem that aims at recovering motion parameters from observed motion blur only. As a result monocular motion estimation is closely related to image deblurring and blind deconvolution. Indeed, since the observed blur is parameterized by the motion, motion parameters can be directly estimated from the blur kernels themselves. Over the last decades a lot of research has been devoted to recovering uniform blur kernels; for motion blur which is not purely translational, however observed blur varies non-uniformly across the image plane, which is why most traditional blind deconvolution techniques are not applicable. In order to recover motion from a single image this work proposes a generative blur model that constructs non-uniform blur kernels according to an affine motion model. By incorporating this model into a variational EM (Expectation Maximization) framework, we are then able to recover the affine parameters by blind deconvolution. As far as the inference is concerned, special care must be taken w.r.t. symmetry because of the directional ambiguity of the motion blur. We finally conduct experiments on ground truth datasets. In particular, we utilize an industrial CNC machine to capture image sequences with high precision. By averaging these capturing sequences, we can synthesize affine non-uniform motion blur, based on which we evaluate the performance of our inference framework.

Typ des Eintrags: Masterarbeit
Erschienen: 2014
Autor(en): Gast, Jochen
Art des Eintrags: Bibliographie
Titel: Estimating Motion from a Single Blurry Image
Sprache: Englisch
Publikationsjahr: 2014
Kurzbeschreibung (Abstract):

Estimating motion from a single image source is a heavily ill-posed problem that aims at recovering motion parameters from observed motion blur only. As a result monocular motion estimation is closely related to image deblurring and blind deconvolution. Indeed, since the observed blur is parameterized by the motion, motion parameters can be directly estimated from the blur kernels themselves. Over the last decades a lot of research has been devoted to recovering uniform blur kernels; for motion blur which is not purely translational, however observed blur varies non-uniformly across the image plane, which is why most traditional blind deconvolution techniques are not applicable. In order to recover motion from a single image this work proposes a generative blur model that constructs non-uniform blur kernels according to an affine motion model. By incorporating this model into a variational EM (Expectation Maximization) framework, we are then able to recover the affine parameters by blind deconvolution. As far as the inference is concerned, special care must be taken w.r.t. symmetry because of the directional ambiguity of the motion blur. We finally conduct experiments on ground truth datasets. In particular, we utilize an industrial CNC machine to capture image sequences with high precision. By averaging these capturing sequences, we can synthesize affine non-uniform motion blur, based on which we evaluate the performance of our inference framework.

Freie Schlagworte: Computer vision, Motion estimation, Optical flow, Motion image analysis, Image deblurring
Zusätzliche Informationen:

89 p.

Fachbereich(e)/-gebiet(e): 20 Fachbereich Informatik
20 Fachbereich Informatik > Graphisch-Interaktive Systeme
Hinterlegungsdatum: 12 Nov 2018 11:16
Letzte Änderung: 12 Nov 2018 11:16
PPN:
Export:
Suche nach Titel in: TUfind oder in Google
Frage zum Eintrag Frage zum Eintrag

Optionen (nur für Redakteure)
Redaktionelle Details anzeigen Redaktionelle Details anzeigen