Representation and

Description

Digital Image Processing

 Representation and Description
• Segmented pixels are usually represented and
described in two forms.
• External Representation
– Boundary (e.g. edges, chain-code)
– Describe using shape features
• Internal Representation
– Region (e.g. connected components)
– Describe using color, texture, or other regional features
• Feature descriptors should be insensitive to size,
translation, and rotation if possible.
 Image shapes

• How do we tell if two objects have the

same shape?
• How can we classify shape?
• How can we describe the shape of an
object?
 Chain Codes
• Represent a boundary by a connected sequence
of line segments of specified length and direction

4-directional 8-directional
 Chain Codes

http://www.cs.unca.edu/~reiser/imaging/chaincode.htmls
 Chain Codes

There are two problems with the definition of the

chain code as given in previous sections:
1.The chain code is dependent on the starting
pixel
2.The chain code is dependent on the orientation
of the object
 Chain Codes Normalization
• Start point normalization
– Treat the chain code as a circular sequence of numbers
– Redefine the starting point so that the resulting sequence of
numbers forms an integer of minimum magnitude

10103322 (arbitrary starting point)

01033221 (after normalization)

• Rotation normalization
– Use the first difference of the chain code

01033221 (chain code)

31330303 (difference code)
Chain Codes
Chain Codes
 Shape Numbers
We now consider the second problem from above:
defining a chain code that is independent of orientation of
the object.
 Shape Numbers
Now suppose we rotate the shape so that it has
a different orientation

Even when normalized, the chain codes are different.

Shape Numbers
Shape Numbers
Shape Numbers
 Polygonal Approximation
• A digital boundary can be approximated with
arbitrary accuracy by a polygon.
• The goal is to capture the essence of the
boundary shape with the fewest possible
polygonal segments.
Minimum Perimeter Polygons
 Merging

Line fit
Splitting
 Signatures
• 1-D functional representation of a boundary
• Examples
– Distance from the centroid to the boundary as a
function of angle
– Tangent angle at each point on the boundary
as a function a angle
• Need to normalized for starting point,
rotation, and scaling
Signatures
 Boundary Segments
• Decompose a boundary into segments

Convex Hull
 Skeletons
• Represent the structural shape of a plane region
• Skeleton is defined via medial axis transform
(MAT)
– Points in a region that have more than one closest
neighbor on the region boundary
 Connected Component Labeling
• Binary image with 0 and 1 values
• Scan left to right, top to bottom
• If the value at p is 0, continue
• If p is 1, examine it’s top and left neighbors,
– If both are 0, assign p a new label, increment label count
– If only one of them is 1, assign that label to p
– If both are 1,
• If they have the same label, assign that label to p
• If they have different label, assign the smaller label to p, and
make a note that the two labels are equivalent (the two
components are connected through p)
• Do a second scan, make equivalent labels the same
Connected Component Labeling

1 1

1 1 1 1 1 2

1 1 1 1 1 1 2 3

1 1 1 1 1 1 1 2 2 2

1 1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1

1 1 4 4

1 1 1 4 4 4

After Labeling (4-connected)

 Boundary Descriptors
• Simple descriptors
– Length
– Eccentricity
– Curvature
• Shape numbers
• Fourier descriptors
 Boundary Length
• Number of boundary pixels as length
– 4-connected: overestimate
– 8-connected: underestimate
• Length from 8-directional chain-code
Length = (vertical and horizontal components) +
2
(diagonal components)
 Eccentricity
• Eccentricity =
major axis length / minor axis length

Major Axis

Basic Rectangle Minor Axis

 Curvature
• Rate of change of slope

Change of Slope
 Shape Numbers
• The first difference of a chain-coded
boundary of smallest magnitude.
• Order of shape number is the number of
digits used to represent the codes.
Shape Numbers
Basic Rectangle
Fourier Descriptors

( x0 , y 0 ), ( x1 , y1 ), , x K 1 , y K 1 

( x k , y k )  x(k ), y (k ) 

s (k )  x(k )  jy (k )
 Fourier Descriptors

1 K 1
  j 2uk 
a (u ) 
K

s 0
s ( k ) exp  K , u  0,1,  , K  1 DFT

K 1
 j 2uk 
s (k )   a (u ) exp   , k  0,1,2, , K  1 Reconstruction
u 0  K  of s(k)

P 1
 j 2uk  Approximation
sˆ(k )   a (u ) exp   , k  0,1,2,  , K  1
u 0  K  to s(k)
 Regional Descriptors
• Simple descriptors
– Area
– Perimeter
– Compactness
– Gray-scale statistics
• Topological descriptors
• Texture
 Simple Descriptors
• Area – number of pixels in the region
• Perimeter – region boundary length
• Compactness – (perimeter)2 / area
– Compactness of a disk = 1
• Gray-scale statistics
– Mean
– Median
– Min/Max
– Variance
 Topological Descriptors
• Topology – study the properties of a figure that
are unaffected by any deformation (or rubber-
sheet distortion)

Hole (H)

Connected Component (C)

• Euler number (E) = C - H

 Euler Number

E = 1-2 = -1
E = 1-1 = 0
 Texture
• Measure smoothness, coarseness, and
regularity.
• Principal approaches
– Statistical
– Structural
– Spectral
Texture
 Co-occurrence Matrix
• For n gray levels, let A be a nxn matrix
• Element aij in A is the number of times that points
with gray level zi occur (in the position specified
by P) relative to points with gray level zj
• P is the position operator
• A divided by the total number is called the co-
occurrence matrix
 Co-occurrence Matrix

0 0 0 1 2 0 1 2
1 1 0 1 1 4 2 1  0
A   2 3 2 1
 
2 2 1 0 0

1 1 0 2 0 0 2 0 2

0 0 1 0 1

P = lower right
 Example Texture Descriptors
• Maximum probability
max cij
i, j
 
• Element difference moment of order k

 i  j k
cij
i j

• Uniformity

 c 2ij
i j

• Entropy

  cij log 2 cij

i j