Image Processing for Physical Data

1
Image Processing for Physical Data

• Xuanxuan Su
• May 31, 2002

2
Outline

• Background
• Physical experimental image data
• Pre-processing method
• Correlation computing
• System Implementation
• Evaluation

3
Image Time-of-Flight Spectrometer
Time-Resolved Images 128 x 128 Pixels 730 Hz
Digital Acquisition 500,000 1,000,000
Frames Mass-Resolved Energies Angular
Distributions
Time-Resolved Waveforms Digital Scope
Acquisition Mass-Resolved Energies Distributions
4
Momentum Image
POLARIZATION AXIS
5
Correlation Approach
Real Space Correlation Images
24 - 15o ( Dj) Angle Sectors 30 - 0.66 eV (
DE) Energy Sectors
6
Image Data

• For each experiment
• 500,000 1,000,000 frames
• 8G 16G uncompressed data
• 5M 150M compressed data
• Pixels are sparse

7
Challenge

• Previous work
• Data compression
• Correlation of sectors
• Low accuracy
• Computational resource is limited
• Large data set cant fit in memory
• More than 3 hours for 600 sectors correlation

8
Applied Technologies

• Clustering
• Correlation
• Sampling

9
Clustering

• Pre-processing method
• Represent a cluster of points by their centroid
• Can be used to achieve data compression

1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
10
K-Means Clustering

• Algorithm
• Randomly choose k cluster centers
• Assign each data to the closest cluster center
• Recompute the cluster centers using the current
  cluster member until a stop criteria is met
• Simply and fast
• Sensitive to initial seed selection

11
Incremental Clustering

• Algorithm
• Assign the first data item to a cluster
• For next data item, either assign it to one
  existing cluster or a new cluster
• Repeat step 2 until all the data items are
  clustered
• Advantage
• Small space requirement
• Non-iterative

12
Correlation Coefficients

• A measure of linear association
• The formula

13
Sampling

• Calculate correlation
• Image sampling
• Problem the number of samples that have a good
  estimation of correlation
• Estimate the accuracy of approximation
• Useful for evaluation
• Pixels sampling

14
System Implementation

• Pre-processing
• Incremental clustering method Spotlize
• Pre-define the radius of clusters

15
System Implementation (cont)

• Progressive Correlation Computing
• Pyramidal grids
• Algorithm
• Compute the correlation in a low resolution, find
  the most correlated grids
• Divide the corresponding grids into smaller grids
• Repeat step 1 2 until a stop criteria is met
• Increase accuracy
• have to multi-scan data set

16
Evaluation

• Run time
• Accuracy of correlation
• Spotlized images vs. original images
• Choose sample pixels

17
Future Work

• Avoid multi-scan data set
• Find a number of sampling images so that the data
  can fit in memory and have high accuracy
• Investigate other association measure methods