Performance study of parallel implementation of texture images using GLCM

Abstract

The process of the creation of texture images derived from a windowed GLCM coupled with the calculation of Haralick features for each window is a time intensive process due to the intense number of calculations involved. This study examines and seeks to quantify the expected increase in processing speed when migrating this algorithm from a traditional serial implementation to a parallel implementation of the same function using modern Graphical Processing Units and the effects of certain parameters such as window size and image size. The components of texture images and some of the factors relating to the efficiency of CUDA code are described. The problem domain was analysed and a serial version of texture window analysis was implemented and checked for accuracy by comparing it to code written in Matlab. The serial code was tested on a 2.4 Ghz Intel core i5 processor while the parallel code was tested on two different GPU cards, a GeForce 310M and a GeForce GTX 620. The final (fastest) implementation used three kernels. Two of these performed gray scale conversion and intensity scaling while the third performed the entire GLCM construction and feature extraction. The results showed that a single large kernel could outperform the collection of small kernels that was used in the alternative implementation. As a result of parallel implementation, texture analysis of a 2048 x 2048 pixels image was found to be up to 44 times faster than the serial version using the GeForce 310M and even faster on the GeForce GTX 620.