Haralick's Texture Features Computed by GPUs for Biological Applications

摘要

This paper presents an approach to speedup the computation of co-occurrence matrices and Haralick texture features, as used for analyzing microscopy images of cells, by general-purpose graphics processing units (GPUs). The sequence of computation steps for the features is analyzed based on a graph and an optimized version of the software is derived. Afterwards, a massive parallel software version for GPUs is designed and implemented. On a single node of a cluster, a speedup of a factor of 360 was obtained compared to the original software version, and a speedup of a factor of 32 was achieved compared to the optimized software version.