Accelerating total variation regularization for matrix-valued images on GPUs

Maryam Moazeni, Alex A. T. Bui, Majid Sarrafzadeh
Computer Science Department, University of California, Los Angeles, USA
Proceedings of the 6th ACM conference on Computing frontiers, CF ’09


   title={Accelerating total variation regularization for matrix-valued images on GPUs},

   author={Moazeni, M. and Bui, A. and Sarrafzadeh, M.},

   booktitle={Proceedings of the 6th ACM conference on Computing frontiers},





Download Download (PDF)   View View   Source Source   



The advent of new matrix-valued magnetic resonance imaging modalities such as Diffusion Tensor Imaging (DTI) requires extensive computational acceleration. Computational acceleration on graphics processing units (GPUs) can make the regularization (denoising) of DTI images attractive in clinical settings, hence improving the quality of DTI images in a broad range of applications. Construction of DTI images consists of direction-specific Magnetic Resonance (MR) measurements. Compared with conventional MR, direction-sensitive acquisition has a lower signal-to-noise ratio (SNR). Therefore, high noise levels often limit DTI imaging. Advanced post-processing of imaging data can improve the quality of estimated tensors. However, the post-processing problem is only made more computationally difficult when considering matrix-valued imaging data. This paper describes the acceleration of a Total Variation regularization method for matrix-valued images, in particular, for DTI images on NVIDIA Quadro FX 5600. The TV regularization of a 3-D image with 128^3 voxels ultimately achieves 266X speedup and requires 1 minute and 30 seconds on the Quadro, while this algorithm on a dual-core CPU completes in more than 3 hours. In this application study we are aimed at analyzing the effective of excessive synchronization, which provides an insight into generally adapting Variational methods to the GPU architecture for other image processing algorithms designed for matrix-valued images.
No votes yet.
Please wait...

* * *

* * *

HGPU group © 2010-2021 hgpu.org

All rights belong to the respective authors

Contact us: