Image Processing Reference
In-Depth Information
Fig. 9.10 Results of re-segmentation. The yellow rectangles are data boundaries. The com-
putation is limited in these local regions to decrease the computing time.
labeled image as the seed of lower level layer (see Figure 9.1).
Let g l be the labels of l th Gaussian pyramid, g i
=
(
g l ,
)
EXPAND
l
i
be the result
of g l expanding l
i times. Higher level label image is interpolated as the same size
of lower layer with NN (nearest neighbor method). That is g l 1 (
j
2 ) ,
i
i
,
j
)=
g l (
2 ,
0
<
N for levels l th and pixel
l
1 th
layer respectively. Thus, EXPAND applied to g l will yield an array of labeled im-
ages. The segmented tissues data can be visualized synchronously with interaction.
If the segmentation supervised by pyramid, acceleration is obvious (see
Table 9.2). Here, two additional layers data (1 and 2 level) are produced as the
course labeled images. Results are obtained through: 2 level labels image is interpo-
lated to 0 level size directly. The pyramid based method introduces differences in-
evitably corresponding to different level data, which are unconspicuous in 2D planar
and volume rendering results (commonly less than 5.2%). In re-segmentation, user-
added new seeds in the local limited rectangle region only cost little time compare
to the whole data process, this merit will benefit for the next operations.
<
(
i
,
j
)
, C l 1 and R l 1 are the width and height of l
Ta b l e 9 . 2 Comparison of CPU and GPU acceleration
Pyramid
Initialization
2Level
Seg.
Labels
Interpolation
0Level
Seg.
Total
time
Speedup
Factor
Method
CPU
4288
13146
2501
1429
21364
38.03
GPU
131
134
37
262
564
1440.6
The boundary processing after segmentation mainly focuses on holes in seeded
region and ragged boundary, a simple solution is to apply erosion and dilation in
the segmented image, but the time cost is a problem. More efficient method is still
worthtoseek.
Fig. 9.11 Results of volume rendering
 
Search WWH ::




Custom Search