Lin Cheng, Hyunsu Cho, and Peter A. Yoon
Image segmentation, Eigenmap, GPU computing
Laplacian eigenmap is one of the most widely used techniques to improve cluster-based segmentation of multivariate images. However, one problem with this approach is its excessive computational requirements, especially when processing large image datasets. In this paper, we aim to employ the emerging commodity graphics hardware of eigenmap-based segmentation. In particular, we present a highly parallel implementation for vessel segmentation using Nvidia's CUDA parallel computing platform. We demonstrate that segmentation steps such as computing the weight matrix can be implemented in a highly parallel fashion. In addition, our approach does not require the computation of the entire spectrum of eigenvalues, which is the most time-consuming step in eigenmap-based segmentation. Instead, we use the Lanczos method to calculate the extreme eigenvalues in parallel. Our experiments based on vessel images of various size achieve a speedup up to 14x over the conventional sequential implementations.
Important Links:
Go Back
