"Performance Measurements of Algorithms in Image Processing"

Binna, Tobias and Hofmann, Markus (2011) "Performance Measurements of Algorithms in Image Processing". Bachelor thesis, HSR Hochschule für Technik Rapperswil.

[thumbnail of thesis.pdf]
Preview
PDF
thesis.pdf - Supplemental Material

Download (12MB) | Preview
  • PDF
    Thesis.pdf - Supplemental Material

Abstract

How much can image processing algorithms be parallelized? In this project we implement
image processing algorithms in a massively parallel manner using NVIDIA
CUDA. Furthermore we analyze the resulting performance gains against current CPU
implementations.
Image processing algorithms are often quite complex and can quickly become very
compute intensive tasks. NVIDIA has evolved a technology called CUDA which is an
extension to the C programming language and provides the opportunity of general
purpose computing on NVIDIA graphics cards, using thousands of concurrent threads.
Thus the use of CUDA in image processing is obvious and the potential performance
bene�ts shall be investigated here in detail.
In this project we implemented two algorithms with CUDA, namely the Hough
transform to extract lines from an image and a template matching algorithm to find a
given pattern in a search image. We analyzed the performance of our implementations
and compared the results with reference implementations from the OpenCV library.
Our results will not only focus on performance issues, but will also give information
about the scalability of algorithms using different graphical processing units (GPU).
Additionally we elaborate on hardware limitations and our experiences made during the
CUDA implementations.

Item Type: Thesis (Bachelor)
Subjects: Area of Application > Image/Video Processing
Technologies > Parallel Computing
Technologies > Parallel Computing > CUDA (Compute Unified Device Architecture)
Brands > nVidia
Metatags > ITA (Institute for Internet Technologies and Applications)
Divisions: Bachelor of Science FHO in Informatik > Bachelor Thesis
Depositing User: HSR Deposit User
Contributors:
Contribution
Name
Email
Thesis advisor
Joller, Josef
UNSPECIFIED
Date Deposited: 24 Jul 2012 07:57
Last Modified: 24 Jul 2012 09:35
URI: https://eprints.ost.ch/id/eprint/149

Actions (login required)

View Item
View Item