An empirical study of parallel solutions for GLCM calculation of diffraction images.

Feature calculation of large amount of images is time consuming. The GPU based CUDA framework offers an affordable solution for calculating image features in parallel. The research focused on an empirical study of different implementations of a general-purpose GPU-based solution for calculating Gray-Level Co-occurrence Matrices (GLCM) and associated features of diffraction images of biological cells. The GLCM features calculated from the diffraction images are used for rapid cell classification with the machine learning algorithm Support Vector Machine (SVM). CPU and GPU versions of the GLCM and textural feature calculations were implemented and evaluated with different configurations. The results of the optimized GPU implementation showed an average speedup of 7 times for GLCM calculation of diffraction images, and average speedup of 9.83 times for feature calculation over the CPU version.