GPU-accelerated 2D OTSU and 2D entropy-based thresholding

Abstract

Image thresholding methods are commonly used to distinguish foreground objects from a background. 2D thresholding methods consider both the value of a pixel and the mean of the pixel’s neighbors, so they are less sensitive to noises than 1D thresholding methods. However, the time complexity increases from $O(\ell ^2)$ to $O(\ell ^4)$, where $\ell$ is the number of gray levels. This paper proposes a parallel algorithm ($O(\ell + \ell \log \ell )$ ) to accelerate both 2D OTSU and 2D entropy-based thresholding on GPU. By dividing the thresholding methods into seven cascaded parallelizable computational steps, our algorithm performs all the computations on GPU and requires no data transfer between GPU memory and main memory. The time complexity analysis explains the theoretical superiority over the state-of-the-art CPU sequential algorithm (O( $\ell ^2)$). Experimental results show that our parallel thresholding runs 50 times faster than the sequential one without loss of accuracy.