"Performance Measurements of Algorithms in Image Processing"

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.