Parallel active contour with Lattice Boltzmann scheme on modern GPU

Solving the level set equation of the geodesic active contour (GAC) model in image segmentation typically requires a large number of complicated calculations. To reduce the computing time, we propose a Lattice Boltzmann (LB) Method based partial differential equation solver for the level set equation. The advantages of the LB method are large time steps (thus less iterations) and easy for parallelization. We derive the formula of LB equation parameters for the GAC model and present an approach of GPU implementation. It is the first GPU implementation of the LB model with re-initialization. We adopt a serious of strategies on GPU memory usage considering the Fermi GPU memory hierarchy to further optimize the performance. Experimental results demonstrated that our parallel LB-GAC method achieves a maximal 500+ speedup over previous serial methods with the same segmentation precision.

1. INTRODUCTION

Active contour is one of the most important tools in the field of image segmentation and understanding. One family of active contour, called geometric deformable models, were first introduced by Caselles et al. (the GAC model) [1] for image segmentation. The main feature is that they do not require any parameter of the boundary curve by minimizing the energy functional which represents the contour's geodesic length. Consequently, the curves can be represented as level sets of higher dimensional functional which handles the curve's topological change naturally. Due to its strong robustness, the original GAC model was extended by utilizing different image features and used extensively in numerous applications in image processing and computer vision.