high performance computing on graphics processing units: hgpu.org

With large number of related images being used for applications such as MR spectroscopy imaging, Object of interest 3D modelling and photo collages, the need of the hour is to accelerate image cosegmentation algorithms. Cosegmentation refers to the process of segmenting common regions from multiple related images. A novel distributed algorithm, CoSand [1], for cosegmentation of large-scale image collections was proposed. CoSand involves preprocessing, solving multiple systems of linear equations and clustering for every image, thereby making it computationally intensive. In this work various approaches to parallelize the MATLAB code for this problem were considered. The most time consuming part of code which is solving multiple systems of linear equations was offloaded to the GPU. Other operations have been performed on multicores. Further, unlike other parallelized image processing algorithms that require a single image data transfer to GPU, CoSand requires a set of image data (500 in our case) accounting for a substantial data transfer overhead. Even in this scenario, the parallel implementation on GPU has achieved a significant performance gain which is comparable to the multicore implementation of the entire application.