Multi-GPU parallelization of a 3D Bayesian CT algorithm and its application on real foam reconstruction with incomplete data set

Nicolas Gac, Alexandre Vabre, Ali Mohammad Djafari
L2S, Laboratoire des Signaux et Systemes (CNRS-SUPELEC-UPS), F-91191 Gif sur Yvette, France
hal-00658653, 10 January 2012




   title={Multi GPU parallelization of 3D bayesian CT algorithm and its application on real foam reonconstruction with incomplete data set},

   author={Gac, Nicolas and Vabre, Alexandre and Mohammad-Djafari, Ali},


   affiliation={Laboratoire des signaux et syst{‘e}mes (L2S) , Laboratoire d’Int{‘e}gration des Syst{‘e}mes et des Technologies – CEA LIST},

   booktitle={Proceedings FVR 2011},


   address={Poitiers, France},

   audience={international },





Download Download (PDF)   View View   Source Source   



A great number of image reconstruction algorithms, based on analytical filtered backprojection, are implemented for X-ray Computed Tomography (CT) [1,2]. The limits of these methods appear when the number of projections is small, and/or not equidistributed around the object. That’s the case in the context of dynamic study of fluids in foams for example, the data set are not complete due to the limited acquistion time. In this specific context, iterative algebraic methods are a solution to this lack of data. A great number of them are mainly based on least square criterion. Recently, we proposed a regularized version based on Bayesian estimation approach. The main problem that appears when using such methods as well as any iterative algebraic methods is the computation time and especially for projection and backprojection steps. In this study, first we show how we implemented some main steps of such algorithms which are the forward projection and backward backprojection steps on multi-GPU hardware, and then we show some results on real application of the 3D tomographic reconstruction of metallic foams from a small number of projections. Through this application, we also show the good quality of results as well as a significant speed up of the computation with GPU implementation (300 acceleration factor).
No votes yet.
Please wait...

* * *

* * *

HGPU group © 2010-2021 hgpu.org

All rights belong to the respective authors

Contact us: