next up previous [pdf]

Next: Accuracy comparison Up: Results Previous: Results

Efficiency comparison

One of the primary means of comparison for the relative effectiveness of these two approaches to image segmentation is the computational efficiency of the method. The following table summarizes the computational expense required to create the examples seen in the paper.


Table 1: Comparison of CPU times for two segmentation methods
    CPU time (s)
Image type Pixels NCIS PRC
Synthetic data 2761000 n/a 31
Field data 55000 156 1

Again, due to memory constraints the existing NCIS implementation is unable to segment an image the size of Figure 1(a). The implementation described here, however, produces an accurate segmentation in 31 seconds; during this time, approximately 55 million edges are created, weighted, and used to segment the graph. The efficiency advantage for the new implementation is quantified using the field data example; in this case, the image is segmented over 150 times faster using the new implementation. These differences are extremely significant and represent a huge savings of time and computational expense, especially for larger problems.

uno-eig uno-segeig
uno-eig,uno-segeig
Figure 8.
Eigenvector (a) used to segment the image in Figure 1(b) according to the NCIS algorithm of Shi and Malik (2000) and adapted for seismic data by Lomask et al. (2007), and the resulting salt boundary (b).
[pdf] [pdf] [png] [png]


next up previous [pdf]

Next: Accuracy comparison Up: Results Previous: Results

2010-05-19