CSE 633 Parallel Connected Component Labeling For Image Analysis Yang Fei

advertisement
CSE 633 2010 Fall
Parallel Connected Component Labeling
For Image Analysis
Yang Fei
Dec. 7. 2010
Computer Science and Engineering
SUNY at Buffalo
Connected Component Labeling
Application
Widely used to group connected Pixels in Image
Distinguish segments of an image
Application: Medical Testing and Meteorological Image
Problem belong to subset of Connected Components
in Graph
Several Parallel problem about this topics
Image Analysis Example
Problem Description
• Using Binary image as example
pixels have two value 1 (object pixels)
0(background pixels)
• Labeling object
pixels with value equal to 1
Under 4-connectivity, adjacent pixels mark
as the same label
Two Pass algorithm
• First Pass
Scan each element of data by row
Assign temporary label to each element
Store equivalence between adjacent pixel
• Second Pass
Based on the equivalence table, scan image
again
Relabel the element with the lowest equivalent
label
Parallel
• Why parallel
time-consuming, need maintain large table for equivalent
label
• Parallel Method
Divide pixels into n parts
Assign each part to different processors
Try to decrease computation time of this algorithm
• Advantages
Each scan only on small size
Don t need maintain large equivalent table
Experiment
• Edge machine
• Used functions
MPI_Scatter MPI_Gather
• Input Image
Simulate binary image as two-dimensional
matrix (two-dimensional array)
Implementation
• Generate two-dimensional matrix on
root processor
• Assign each part to other processor by
MPI_Scatter
• Run two- pass algorithm on each processor
• Collect label information from each processor
• Relabel adjacent pixels on boundary and
relative pixels
Result
30
8 processors
25
4 processors
Time(ms)
20
15
10
5
0
64
256
576
1024
1600
2304
Number of Pixels
3136
4096
5184
6400
Result
40
35
30
time(ms)
25
20
pixels=6400
15
10
5
0
4
5
6
7
8
9
10
number of processors
12
14
16
Reference
Susanne Hambrusch
A Study of Connected Component Labeling Algorithm on The MMP
D. S. Hirschberg
A Parallel Graph Algorithm For Finding Connected Components
Jung-Me Park, Carl G.Looney
Fast Connected Component Labeling Algorithm Using A Divide and
Conquer Algorithm
Download