Image segmentation with local active contours on graphics processing units

Active contours model (ACM) is one of the most attractive and popular methods for image segmentation. While the different formulations of the ACM are generally known to be highly computationally demanding, like most of the image processing techniques do, they are however potential candidates for parallelization.

The graphics processing units (GPUs) are nowadays increasingly present in all kind of image processing among other research areas. In this paper, our goal was the implementation of a GPU parallelized local based formulation of the active contours model, namely the local binary fitting energy (LBF) algorithm. This algorithm involves a large amount of convolution operation computations. Besides the convolution, many other finite differences based operations involved in the algorithm are readily parallelized. So, almost all the heavy computational work is moved to the GPU side, and the CPU needs only to control the main loop and some intermediate computations. To validate our results, we have been experimenting with medical images of different sizes, and different convolution filtering sizes, by comparing the speedup of the GPU parallel implementation vs the CPU sequential implementation.