HPM A Humanlike Automatic Landmarking Algorithm

1
HPM A Humanlike Automatic Landmarking Algorithm

• Sam Chen
• 11/19/08

2
Seminar

• Seminar is, generally, a form of academic
  instruction, either at a university or offered by
  a commercial or professional organization. It has
  the function of bringing together small groups
  for recurring meetings, focusing each time on
  some particular subject, in which everyone
  present is requested to actively participate.
  This is often accomplished through an ongoing
  Socratic dialogue with a seminar leader or
  instructor, or through a more formal presentation
  of research.

3
Why Need Landmarking

• To provide feedback information on INR accuracy
  to control system
• To verify if the system meets INR requirements

4
Why Humanlike?

• Human eye/brain system is the best image
  processing system. It automatically adjusts to
  changes in intensity, color, texture to identify
  land and sea. It uses global information instead
  of local only. In practice, when RPM
  autolandmarking fails to work, human operator
  sometimes can still perform landmarking
  successfully

5
How to Landmark?

• By aligning a template with certain known
  features such as land/sea boundary, bright spots
  etc. to their corresponding party in the image

6
How to align?

• Usually use a maximization process by moving the
  landmark template to the location such that a
  function defined by the template and the
  corresponding image part (meaning under the
  template) takes maximum value, i.e., they match
  each other.

7
What Functions to Use?

• Because we are matching two features (in digital
  world two vectors) we can use normalized cross
  correlation or cross covariance
• The definitions are slightly different from those
  in statistics

8
Definition of Normalized Cross Correlation

• For two vectors u and v, their (non-centered)
  normalized cross correlation is the inner product
  of u and v divided by the L2 norm of u and the L2
  norm of v
• Geometrically, it is the cosine of the angle
  between the two vectors when the two vectors
  are parallel, their normalized cross correlation
  1.

9
Definition of Cross Covariance

• Only the inner product in the numerator

10
Example of Correlation Maximization matching

• Assume we have a landmark template
• 1 2 and an image 4 8 8 8
• 1 1 4 8 8 8
• 4 4 8 8
• 4 4 8 8
• where is the match?
• 4(1 2 1 1) (4 8 4 4)

11
Underlying Assumption for Image Correlation
Maximization

• The feature image does not change over time
• 1 2 will not match 2 3 3 3
• 1 1 2 3 3 3
• 2 2 3 3
• 2 2 3 3
• exactly (normalized cross correlation will
  never be 1)

12
Advantage and Disadvantage of Normalized
Correlation Maximization

• It is a real pattern matching the two vectors
  must be close to parallel. Bright spots will
  reduce the match
• Landmark template must from real image but
  illumination may change by season and time of the
  day
• If we only have map based landmark templates,
  they will not correlate to image well leading to
  wrong location

13
Underlying Assumption for Cross Covariance
Maximization

• Coastline has the highest overall contrast

14
Advantage and Disadvantage of Cross Covariance
Maximization

• It can use segments of coastline from map without
  intensity as templates
• Because it maximizes the sum of intensity, edge
  of bright spots becomes attractors. In order to
  remove such bright spots, elaborate cloud
  detection algorithm needs to be introduced.
  However, that algorithm needs good attitude.
  Which comes first? Chicken or Egg?

15
What are The Key Considerations of This Algorithm

• Human eyes examines the whole image to get an
  idea what looks like land and what looks like sea
  before determining locally the land/sea boundary
• There are three and only three different pixels
  land, sea, cloud (can not see through)
• Human defines a cloud pixel as one which blocks
  the view of land or sea. If the pixel looks like
  a land pixel, it is a land pixel. Human PM
  operator does not care about scientific
  definition of cloud
• Coastline is long and connected

16
Histogram Clustering

• Calculate histogram of neighborhood image
• Find peaks and valleys of the histogram
• Three clusters only
• Set pixels between two valleys to intensity at
  the peaks
• All pixels beyond the second valley are to be
  treated as cloud pixels

17
Find Coastline (Edge Detection)

• Differentiate the three-intensity image using
  Sobel Operator to get a Sobel image
• Set all cloud pixels and their immediate
  neighbors to intensity zero
• Using a threshold to make the Sobel image a 0/1
  image
• Clean up the image (to be discussed later)

18
Correlation

• Correlate digital chip (a 0/1 image to be defined
  later) with the 0/1 Sobel image chip
  image
• 0 0 1 0 0 0 0 1 0
• 0 0 1 0 0 0 0 1 0
• 1 1 1 0 0 1 1 0 1
• 0 1 0 0 0 0
• 0 1 0 0 0 0

19
Definition of Digital Chip

• A landmark template of coastline from a map is a
  set of vectors connecting base points
• The template is laid on top of an image according
  to navigation information
• Set any pixel containing one or more points from
  the template as 1 and set the rest to zero
• This creates a digital chip

20
Subpixel Accuracy

• After correlating digital chip to 0/1 Sobel
  image, one can only get accuracy at line/pixel
  level
• We need to overcome digitization error when
  generating digital chip

21
Interpolation NOT Recommended

• The image data is not smooth
• The Sobel image, which is the derivative of a
  discontinuous function with sharp peaks, is not
  suitable for interpolation

22
Perturbation to The RescueAn Example

• Suppose that you have a digital voltmeter with a
  resolution of 1 v
• You want to measure a DC voltage of 10.3 v
  accurately
• You add a sequence of perturbations to the
  voltage you want to measure and get a sequence of
  results 10, 10, 11,10, 10,
• You average the results

23
Perturbation in HPM

• After laying landmark template to image according
  to navigation information, perturb the location
  by fractional line and pixel
• Use the same way to create a digital chip which
  may or may not be the same as the original one
• The number of different chips are small (compare
  to the number of perturbations)

24
Perturbation in HPM

• Calculate correlation for each different digital
  chip
• Average over the number of perturbations to get
  subpixel accuracy
• Perturbation can be a uniformly distributed
  random number from -1 to 1 or deterministic -1,
  -0.9, -0.8,

25
Some test Results

• Original image
• Sobel image
• Histogram
• Segmented image
• Sobel of segmented image

26
Original image
27
Sobel Image
28
Histogram
29
DefuziedImage
30
Sobel from DefuziedImage
31
Whats Next

• Test on images with more clouds and those current
  RPM will misidentify
• Test other channels
• Improve histogram segmentation algorithm
• Write codes to eliminate small closed curves on
  the right side using topological methods.