HIGH LEVEL PROCESSING

1

• HIGH LEVEL PROCESSING
• After you have improved the quality of an image,
  and identified object boundaries, you are ready
  to embark on your ultimate objective of object
  recognition.
• This process is not a one step process. you still
  have to proceed to some more processing steps --
  known as high level processing -- before
  considering object identification.
• One simple way to identify the shape of an object
  is to generate information about edges. Another
  way which can be very effective in some
  applications is to work with silhouettes.
• This is not fundamentally different from an edge
  description - here an assumption is made that
  overall shape

2

• HIGH LEVEL PROCESSING
• characteristics - rather that fine detail or
  subtle shade variations - are of principal
  importance.
• Typical characteristic could be surface area or
  perimeter length.
• This method assumes that an object is defined by
  roughly uniform surface rather than
  object/background discontinuities at edges or
  other surface locations.
• This type of segmentation can be easily
  accomplished by applying a thresholding
  procedure to the raw gray scale image.
• This works well for simple images, which are
  characterized by shapes (uniform lighting) and no
  shading.

3

• HIGH LEVEL PROCESSING
• Contour Tracing ( for edge information inside and
  on the boundary)
• There may be situations where there is a need to
  generate edge information (not just boundary),
  which can be processed by computer (edge
  description - length, direction, shape, contour
  etc)- sometimes holes in an image.
• We have to develop a procedure that will generate
  edge description from a silhouette or region
  based description.
• The principle of this approach is to find any
  point on the edge of an object and then by
  utilizing local information, track around
  successive points on the edge (marking each point
  traversed until the starting point is reached.
• ? ? ? ? ? ?
• ? ?
• ? ? ? ? ? ?
• The procedure is simple.

4
p
5
Apply contour trace
Outline obtained from contour tracing algorithm
Silhouette of an Image
6

• HIGH LEVEL PROCESSING
• Contour Tracing

Start
Locate a point Pi the object boundary
Turn through 90 deg counter clockwise
Check New pixel in new direction
Turn through 90 deg clockwise
Yes
Is Pi regained ?
STOP
No
Is new point inside boundary
No
Yes
Mark it
7

• HIGH LEVEL PROCESSING
• Region Oriented Segmentation
• The objective of the segmentation is to partition
  an image into regions based on different
  intensities.
• The problem is approached by finding boundaries
  between regions based on intensity
  discontinuities.
• We perform segmentation of an image into distinct
  regions (within each region, pixel points are
  assumed to be connected in some way).
• Let R represent the entire region. The
  segmentation process partitions the region R into
  R1, R2, Rn, such that each region Ri is uniform
  ( connected) in terms on intensity values.
• By this process, we are able to identify
  individual objects by merging the regions having
  similar intensity values.
• The simple process can be described as
• A first pass scans the gray scale, raw image and
  identifies primary regions by connecting together
  group of pixel- points of constant gray scale
  intensity values.

8
(No Transcript)
9

• HIGH LEVEL PROCESSING
• Region Segmentation
• A second pass through the image now works with
  these primary regions and examines boundaries
  between these regions already identified.
• Depending on differences in intensity values
  between primary regions, these regions may be
  merged to give a wider ranging regions or may be
  left as isolated pixel groups. ( see figure next
  page )
• In this case the segmentation had identified what
  might easily be interpreted as two objects - a
  large object with a smaller object by its side.

10
(No Transcript)
11
A practical image using region segmentation
approach is shown below
12

• HIGH LEVEL PROCESSING
• Object Representation and Coding
• Since an object of interest in an NXN image may
  occupy small number of pixels - ( may be
  typically 32x32 or 64x64 pixels) it would be
  really wasteful to process the entire image.
• As successive manipulations on an image require
  time consuming and complex computations.
  Therefore, it is useful to consider more
  appropriate means of encoding the image data.
• There exist a very useful technique for object
  representation - the chain coding - it gives the
  object of interest.
• Chain encoding employs the principle of finding
  an arbitrary initial point on the image boundary.
  This point is given spatial coordinates. This
  starting point is useful for subsequent image
  encoding.

13

• HIGH LEVEL PROCESSING
• Object Representation and Coding
• Subsequently, only the directional changes are
  recorded and stored. This procedure is continued
  along the edges until the original point
  (starting point) is regained.
• To record directional changes following
  directions are defined.

14

• HIGH LEVEL PROCESSING
• Object Representation and Coding
• See figure on the next transparency - with
  starting point.
• The chain code is defined as as follows
• 1 1 0 0 0 0 6 0 6 6 6 4 6 4 4 4 4 3 3 2 ( derived
  from the table on the next page).
• It is very clear that such coding scheme can
  significantly reduce the storage requirements for
  many images of practical interest.
• The chain code technique has an advantage beyond
  these storage considerations. For example, it may
  be necessary to obtain measurements of
  identifying features that describe some physical
  characteristics of an object.
• For example we may wish to measure, the area of
  an object or perimeter. This chain coding allows
  computation of some basic physical properties
  directly without the necessity to reconvert it
  into NXN pixel format.

15
(No Transcript)
16
(No Transcript)
17

• HIGH LEVEL PROCESSING
• Object Representation and Coding
• Some of the physical parameters are defined as
  perimeter, length, width and area.
• Perimeter length of codes 0,2,4,6 is 1 and that
  of 1,3,5,7 is ?2.
• The total perimeter of an object P SE ?2SO
  16 ?2(4) 16 5.66 21.66 units.
• Area Measurement object area is obtained by
  direct computation from the chain code,
  evaluating the elemental contributions of several
  vertical strips bounded by the current
  directional transition path between two
  successive edge pixels and the x-axis.
• Increments are additive, subtractive or of zero
  contributions according to following rules.

18
Neutral

Additive
Subtractive
19
(No Transcript)
20

• HIGH LEVEL PROCESSING
• Object Representation and Coding

2
1
3
0
4
5
7
6
21

• HIGH LEVEL PROCESSING
• Object Representation and Coding
• A step in 0 direction adds 1 y units of area.
• A step in 5 direction would be subtractive
  1(y-0.5).
• A step in direction 2, which follows a vertical
  path, contributes a zero component to the area
  calculation.
• Applying these principles to the complete set of
  possible directions generates the table of
  contributions.
• The total area of the figure shown will be
• A 5.5 6.5 7 7 7 7 0 6 0 0 0 - 3
  0 - 2 -2 - 2 - 2 - 2.5 - 3.5 0 29 square
  units.
• This can be checked by the reference diagram.

22
(No Transcript)
23

• HIGH LEVEL PROCESSING
• Object Width
• The maximum width of an object is measured across
  the horizontal line, and can be obtained by
  similar principle, by counting incremental steps
  in x direction (ve or ve only)
• For the given object - width is positive and
  is 7 pixel length.
• Similarly vertical shift or height can also be
  determined.
• Differential Chain code
• It is generated using the chain code by taking
  the difference in the direction ( counter
  clockwise ) of two subsequent chain codes. This
  is useful in getting rid of the rotational
  effects in an image.
• Example get the differential chain code for the
  image
• 1 0 3 3 2 1 0 Verify the DCC as 1 3 3
  0 3 3 3 .

24
(No Transcript)
25

• HIGH LEVEL PROCESSING
• Noise Reduction Algorithm

26
start
Get Pth pixel
Let N1
yes
Is value of P 0 ?
No
No
Value of Nth Neighbor 0
i
N N1
yes
no
Is N 8 ?
yes
Change P 0
27

• Feature Selection For Object Recognition
• For the recognition of an object, you would
  require certain features from the image under
  test and try to match them with the reference
  image features.
• These features must be simple and not very
  computational intensive.
• Following are a few of the possible features.
  This list of features is by no means complete.
  These are a few of the common features.
• Texture Based Features
• A high pass filter shows high frequency
  components. A lot of high frequency components
  implies a course texture. If the image does not
  contain high frequency, then the texture is
  smooth.
• Fourier transform operation can also be used ,
  which directly

28

• gives high frequency components.
• Sometimes it is possible to look at the image
  brightness variations and get an idea about high
  frequency components.
• If there are high frequency components, image
  contains brightness variations.
• Brightness Value and Color Features
• For the brightness value in an image, the
  histogram is a good place to start. Histogram
  shows brightness distribution of an image.
• For color objects, RGB component pixel values of
  an object can be converted to HSI color space.
  Then, examination of the histogram of the hue
  component of the image will show predominant
  color values of the object.
• For color sorting based applications, this
  feature alone will be enough to classify an
  object correctly.

29

• Statistical values of brightness of an object can
  also be very useful feature measures. The mean
  brightness with standard deviation can give good
  features for an object.
• Mode brightness is the most common brightness
  feature used for object recognition. This feature
  consists of sum of all image pixel brightness.
  This corresponds to energy content of an object
  This feature is called zero-order spatial moment.
  
• These features are very useful in object
  recognition problem. However, it should be noted
  that a single feature is never enough for a
  reliable object identification. As a rule of
  thumb, a combination of minimum of 2-3 features
  is suggested.
• Shape Measure
• Shape measurements reflect physical dimensional
  measures that can characterize the appearance of
  an object. As such the list of shape measures can
  be very long.

30

• Any particular image analysis application will
  make use of a few of the shape features (
  descriptors).
• Some of these measures are
• Perimeter The number of pixels around the
  circumference of the object. You have to use the
  distance concept to measure the perimeter.
  Euclidean distance is normally used to calculate
  this feature.
• Area Represents the pixel area of the interior
  of the object. The area is computed as the total
  number of pixels inside and including the object
  boundary. The result is a measure of the size of
  the object.
• This measure normally excludes the holes inside
  the object boundary.
• Area to Perimeter Ratio A ratio based on the
  area and perimeter2 . It is is an excellent
  measure. The result is a measure of object
  roundness or compactness.

31

• This value is between 0 --1. Greater the ratio,
  rounder the object.
• R 4 ? x area / perimeter 2
• Major Axis The (x,y) end points of the longest
  line that can be drawn inside ( through ) the
  object.
• The major axis endpoints (x1,y1) and (x2,y2) are
  found by computing the pixel distance between
  every combination of border pixel points and
  finding the pair with longest distance.
• Major Axis Length This is the pixel distance
  between the major axis endpoints. It is defined
  as
• length SQRT( x2-x1) 2 ( y2-y1) 2 )
• Major Axis Angle It is the angle between the
  major axis and horizontal axis. This can be
  from 0 0 - 360 0 .
• angle tan -1 ( y2 - y1)/( x2-x1)
• where (x1,y1) and (x2,y2) are major axis endpoints

(x2,y2)
?
(x1,y1)
32

• Minor Axis Minor axis is perpendicular to the
  major axis. The minor axis endpoints (x3,y3) and
  (x4,y4) are found by computing the pixel distance
  between two border pixel endpoints.
• If there is a choice, the largest distance must
  be selected.
• (x3,y3) and (x4,y4) are the end points on the
  boundary of the image.
• Minor Axis Width It is the distance between
  minor axis endpoints. The result is a measure of
  object width.
• minor axis width SQRT ( x4-x3) 2 ( y4 -
  y3) 2
• Ratio of Minor Axis Width to Major Axis Length
  This ratio is

(x4,y4)
horizontal axis
(x3,y3)
33

• computed as the minor axis width distance
  divided by the major axis length. The result is a
  measure measure of elongation. The value of this
  ratio is between 0 and 1.
• Boundary Box area the smallest area that would
  surround the object. It is defined by a rectangle
  formed out of minor axis width and major axis
  length.
• BBA major axis length x minor axis width
• Other features could include
• Number of holes inside the object
• Total number of connected objects
• Center of gravity
• Chain codes ( commonly used) and DCCs
• Fourier descriptors ( very commonly used)
• Object Signature ( very commonly used )

34

• It is also suggested to use relative values of
  the angles and the normalized distances. This
  makes these features as invariant to scaling and
  rotations ( ratios are ideal feature measures).