Implementation and Performance of Image Filtering on GPU

摘要

Image filtering is one of the most important parts in the image-processing. It takes much more time to performance the convolution in image filtering on CPU since the computation demanding of image filtering is massive. Contrast to CPU, GPU may be a good way to accelerate the image filtering. CUDA (Compute Unified Device Architecture) is a parallel computing architecture developed by NVIDIA. CUDA is highly suited for general purpose programming on GPU which is a programming interface to use the parallel architecture for general purpose computing. This interface is a set of library functions which can be coded as an extension of C language. In this paper, the filtering was implemented in the frequency domain instead of the spatial domain since the filtering in the frequency domain is faster than convolution in the spatial domain if we have filters with many coefficients and filtering in 2D. Compared with the traditional method of image filtering which was performed on CPU, the implementation of image filtering on GPU has a speedup of approximately 10 times. GPU has great potential as high-performance co-processor.