Conjugate gradient in noisy photometric stereo

Ryszard Kozera, Felicja Okulicka-Dluzewska

Research output: Chapter in Book/Conference paperConference paper

2 Citations (Scopus)

Abstract

This paper discusses the problem of reconstructing the Lambertian surface from noisy three-light source Photometric Stereo. In the continuous image setting the shape recovery process is divided into two steps: an algebraic one (gradient computation) and analytical one (gradient integration). The digitized case with added noise has it discrete analogue in which also perturbed gradient from three noisy images is first computed. Generically such non-integrable vector field is subsequently rectified to the ”closest” integrable one. Finally, numerical integration scheme yields the unknown surface. The process of vector field rectification is reduced to the corresponding linear optimization task of very high dimension (comparable with the image resolution). Standard methods based on matrix pseudo-inversion suffer from heavy computation due to the necessity of large matrix inversion. A possible alternative is to set up an iterative scheme based on local snapshots’ optimizations (e.g. 2D-Leap-Frog). Another approach which is proposed in this paper is solving the above global optimization scheme by Conjugate Gradient with no inversion of matrices of large dimension. The experimental results from this paper show that the application of Conjugate Gradient forms a computationally feasible alternative in denoising Photometric Stereo. © 2014 Springer International Publishing Switzerland.
Original languageEnglish
Title of host publicationICCVG 2014: Computer Vision and Graphics
EditorsLeszek J. Chmielewski, Ryszard Kozera, Bok-Suk Shin, Konrad Wojciechowski
PublisherSpringer
Pages338-346
Number of pages9
ISBN (Electronic)978-3-319-11331-9
ISBN (Print)978-3-319-11330-2
DOIs
Publication statusPublished - 2014
Externally publishedYes
EventInternational Conference on Computer Vision and Graphics ICCVG 2014 - Warsaw, Poland
Duration: 15 Sep 201417 Sep 2014

Publication series

NameLecture Notes in Computer Science
Volume8671
ISSN (Print)0302-9743

Conference

ConferenceInternational Conference on Computer Vision and Graphics ICCVG 2014
CountryPoland
CityWarsaw
Period15/09/1417/09/14

Fingerprint Dive into the research topics of 'Conjugate gradient in noisy photometric stereo'. Together they form a unique fingerprint.

Cite this