PARALLELIZATION OF IMAGE PROCESSING AND COMPUTER VISION RELATED ALGORITHMS USING COMMODITY NVIDIA GPGPU

Abstract

In this project we used NVIDIA commodity GPU to parallelize image processing algorithms for speedup of part of the post-processing of data from a 3D imaging device so that the post-processing time could be practically acceptable. CUDA (short for œCompute Unified Device Architecture) introduced by NVIDIA was used to implement the algorithms on GPU. Much work has been done on optimizing GPU programs mainly focusing on memory and execution configurations optimizations. The image processing algorithms we worked on include denoising, nearest neighbor search, contrast enhancement and image segmentation. Performance gain of 18x and 10x was achieved for denoising and nearest neighbor search respectively, which is up to our expectation. However only 3x speedup was achieved for contrast enhancement. By examining this problem we realized that not all problems would be suitable for GPU parallelization.