ABSTRACT

Digital image processing involves operations like image enhancement, image restoration and image compression, just to name a few. The representation of a digital 2- dimensional image in computer is visualized a s 2-D arry. That is why th operations performed on these images are treated as if they were being applied on 2-D arrays. Traditional serial execution of these image processing operations in CPU is costly in items of execution time since there is only compute engine and a lot of data to process. These operations can be speed up to a noticeable extent if parallel processing techniques are applied in a well way. The NVIDIA compute unified device
Architecture (CUDA) provides an Application Programming Interface (API) to writ te code which is partially parallel and can be executed on a Compatible Graph ics Processing Unit (GPU) in collaboration with the CPU. The serial part of the code executes on CPU and the parallel par1 of the code is seamlessly ported to the GPU. There are mainly two approaches for digital image filtering. In the spatial domain, the image is filtered by applying the filter mask directly on the
Image pixels. In the frequency domain, the image is first transformed and then filtered fol lowed by transforming it back to the spatial domain. Since the digital image processing involves a lot of data processing and the operation being applied on all data elements (pixels) is most of the time same, they are the ideal candidates for the data level parallelism. This kind of parallelism can be implemented using CUDA
with a compatible GPU since N VIDIA GPUs now available with hundreds of CPU cores and a significant amount of on chip memory.
There are many noticeable research works that have been carried out to take advantage of CUDA in image processing. Thus the aim of this research is to find out the application of CUDA in the image filtering in frequency domain and to
implement the Butterwort h Low pass filter in CUDA . Finally, the speed t'.p of the
CUDA version of Butter Low pass filter is compared with its CPU-only version.