Massive Parallel Image Processing

Massive Parallel Image Processing: How can image data be processed in an extreme
parallel manner. In this project, we analyze methods on how image segmentation could
be developed with CUDA and give an overview of the advantages and disadvantages by
using CUDA in image processing.
Image processing algorithms are more often than not quite complex and a special
part of them - the image segmentation task - can become quickly very long-winded
because each pixel has to be analyzed and processed repeatedly. NVIDIA has provided
a technology called CUDA, based on the C programming language that supports
calculations on their graphics cards with thousands of concurrent threads. For this
reason the use of CUDA to solve image segmentation algorithm problems is obvious
and the applicability of CUDA in this area should be investigated.
We have developed an application that implements an automatic seeded region
growing algorithm which divides a given image into color based regions, using the
power of NVIDIA's graphics processing unit on most of the partial sub algorithms. The
application delivers an output where the regions in the resulting image are colorized
with their color mean value additionally to a console output, showing the time required
for each algorithm part. The application can be launched with diff�erent command line
arguments which provides the ability to observe, how the results diff�er while playing
with various threshold values.
Equally important to the implementation documentation we elaborate on the lessons
learned from the challenges and performance insights. In addition, we deliver information
about the expedient use of the CUDA technology and what de�finitely should be avoided.