Image segmentation based on Normalized Cut and CUDA parallel implementation

This paper proposes a parallel algorithm using CUDA GPU to accelerate the process of image segmentation algorithm based on Normalized Cut. After giving a summary of the key concepts and theory of normalized cut and CUDA, detailed implementation issues are discussed including the calculation of affinity matrix, transforming symmetric matrices to symmetric tridiagonal matrices, calculation of generalized eigenvalue value and its associated eigenvetor, the choice of splitting point, stopping criterion etc. This algorithm doesn´t sparse the similarity matrix, so there is no information loss in transforming, which will lead to a more real and reliable segmentation. The experiment shows that the parallel algorithm using CUDA not only segment the image reliably but also have a great performance speed-up.