Parallel Feature Extraction - High-Throughput Image Reconstruction and Analysis - page 152

Figure 6.4 In a contour-centric model, each process is responsible for handling the entire contour,
and acquires the relevant image data from successive slices.

6.5 Computational Results

We present preliminary results obtained by implementing some of the image pro-
cessing algorithms described above on parallel architectures. This constitutes re-
search work that is in progress, and significant effort remains towards building a
general-purpose parallel image processing library. These results show the signifi-
cant computational speedup that can be achieved through parallelization.

6.5.1 Median Filtering
We implemented the iterated median filter operation on the Blue Gene/L supercom-
puter, using MPI. We found that a 5

× 5 × 5 filtering operation takes 38 seconds

for 64 slices of 2,048 8

cartesian grid. The same computation running in MATLAB on a single Intel
2-GHz processor took 143 minutes, or about 250 times slower. This illustrates
that the use of high-performance computing dramatically improves the through-
put of image processing tasks. Figure 6.8 depicts the variation in processing time
with the number of processors used. Further details of the processing times on
different architectures are provided in [30].

× 1,872 grayscale images on 512 processors in an 8 × 8 ×

Figure 6.5 Image data from five successive slices represented by a z index ranging from 1 to 5.
This image data constitutes the input to the 3D median filter.