Image Segmentation

1
Image Segmentation
Dr. Ramprasad Bala Computer and Information
Science UMASS Dartmouth CIS 465 Topics in
Computer Vision
2
Image Segmentation

• The term image segmentation refers to the
  partition of an image into a set of regions that
  covers it.
• The goal in many tasks is for the regions to
  represent some meaningful area.
• Regions may also be defined as group of pixels
  having both a border and a particular shape such
  as a circle, ellipse or polygon.

3
Image Segmentation (IS)

• Segmentation has two objectives
• The first is to decompose the image into parts
  for further analysis.
• The second is to perform a change in
  representation.

4
First Objective

• In simple cases, the environment might be well
  enough controlled so that the segmentation
  process reliably extracts only the parts that
  needs to be analyzed.
• In more complex cases, a great deal of domain
  knowledge may be required. Such as extracting a
  complete road network from a grayscale image.

5
Second Objective

• The pixels of the image of the image must be
  organized into higher-level units that are either
  more meaningful or more efficient for further
  analysis.
• A crucial issue is the development of a bottom-up
  approach that is domain independent (usually this
  is very difficult).

6
(No Transcript)
7
Identifying Regions

• Regions of an IS should be uniform and
  homogeneous with respect to some characteristic -
  gray-level, color or texture.
• Region interiors should be simple and without
  many holes.
• Adjacent regions of a segmentation should have
  significantly different values with respect to
  the characteristic on which they are uniform.
• Boundaries of each segment should be smooth, not
  ragged, and should be spatially accurate.

8
Clustering Methods

• Clustering in Pattern Recognition is the process
  of partitioning a set of pattern vectors into
  subsets called clusters.
• The general problem in clustering is to partition
  a set of vectors having similar values.
• For example, consider pattern vectors of two real
  numbers. Then clustering could consist of finding
  subsets of points that are close to each other
  in Euclidean two-space.

9
Classical Clustering Algorithms

• In image analysis, the vectors represent pixels
  or sometimes small neighborhoods around pixels.
• The components of these vectors could be
• Intensity values
• RGB values and color properties derived from them
• Calculated properties
• Texture measurements.
• Any feature that can be associated with a pixel
  can be used to group them.

10
Iterative K-Means Clustering
11
Iterative K-Means Clustering

• This algorithm is guaranteed to terminate, but it
  may not find the global optimum.
• Step 2 may be modified to partition the set of
  vectors into K random clusters and then compute
  their means.
• Step 5 can be modified to stop after the
  percentage of vectors that change clusters in a
  given iteration is small.

12
(No Transcript)
13
Isodata Clustering

• Isodata clustering is another iterative algorithm
  that uses a split-and-merge technique.
• Assume that there are K clusters C1, C2, ,Ck
  with means m1, m2,,mk and Sk be the covariance
  matrix of cluster k.

14
(No Transcript)
15
Isodata clustering algorithm
16
(No Transcript)
17
Histogram-Based Method

• Intuitively Histograms represent regions of
  common features gray-scale or color.
• Finding valleys and segmenting based on peaks can
  be a simple solution.
• Finding valleys automatically is a non-trivial
  problem. (Histogram mode seeking).
• May need further processing using domain
  knowledge.

18
(No Transcript)
19
(No Transcript)
20
Ohlanders Recursive Algorithm

• Ohlander, Price and Reddy (1978) proposed a
  recursive approach to the histogram-based
  clustering.
• The idea is to perform histogram mode seeking
  first on the whole image and then on each of the
  regions obtained from the resultant clusters,
  until regions are obtained that cannot be
  decomposed any further.

21
(No Transcript)
22
Region Growing

• Instead of partitioning the image, a region
  grower begins at one position and attempts to
  grow each region until the pixels being compared
  are too dissimilar to the region to add them.
• Usually, a statistical test is performed to
  decide if it is the case.

23
Haralicks region growing

• Assumes a region is a set of connected pixels
  with the same population mean and variance.
• Let R be a region of N pixels neighboring a pixel
  with gray-level intensity y.
• Let X and S2 be respectively the mean and the
  variance of the region R.
• The statistic T is computed as follows

24

• If T is small then y is added to the region and
  the new mean and variance updated.
• If T is too high y is not likely to have risen
  from the population of pixels in R.
• To give a precise meaning to the notion of too
  high a difference, we can use an a-level
  statistical significance test.
• The fraction a represents the probability that a
  T statistic with N-1 degrees of freedom will
  exceed tN-1(a).
• If the observed T is larger than tN-1(a) then we
  declare the difference to be significant.

25
If the pixel y and segment really came from the
same population, the probability that the test
provides an incorrect answer is a . a is a user
specified parameter. tN-1(a) is higher for small
degrees of freedom and lower for larger degrees
of freedom.
26
Representing Regions

• Each algorithm that produces a set of image
  regions has to have a way to store them.
• Some of the methods are -
• Overlays on the original image
• Labeled images
• Boundary encoding
• Quad-tree structures
• Property tables

27
Overlays

• An overlay is a method of showing the regions
  computed from an image by overlaying some color
  or colors on top of the original image.
• To show region segmentation, one could convert
  the pixels of the region borders to white and
  display the transformed gray-tone image.
  Sometimes more than one-pixel width borders are
  used.

28
(No Transcript)
29
Labeled Images

• Labeled images are good intermediate
  representation for regions that can also be used
  in further processing.
• The idea is to assign each detected region a
  unique identifier (an integer) and create image
  where all the pixels of a region will have a
  unique identifier as pixel.

30
(No Transcript)
31
Boundary Coding

• Regions can also be represented by their
  boundaries in a data structure instead of an
  image.
• The simplest form is just a linear list of border
  pixels.
• A variation of the list of points is the Freeman
  chain code, which codes the information from the
  list of points at any desired quantization and
  uses less space than the original point list.

32
(No Transcript)
33
Quadtrees

• An image is broken into regions.
• Each region of interest would be represented by a
  quadtree structure.
• Each node of a quadtree represents a square
  region in the image and can one of three labels
  full, empty or mixed.
• If the region is labeled full then every pixel in
  the region is of interest.
• Only mixed regions have children.

34
(No Transcript)
35
Property Tables

• Sometimes we want to represent a region by its
  extracted properties rather than by its pixels.
  This is called a property table.
• It is a table in a relational database sense that
  has a row for each region in the image and a
  column for each property of interest.
• Properties can be size, shape, intensity, color
  or texture of the region. Area, ratio of
  minor-to-major axis of the best-fitting ellipse,
  two main colors and more texture measures can all
  be used.

36
Identifying Contours

• Tracking Existing Region Boundaries
• Once regions have been segmented, the border
  pixels for each region can be extracted.
• The border will be a set of pixels that envelopes
  a region.
• The algorithm is quite straight forward.
• Single pass would do.

37
(No Transcript)
38
(No Transcript)
39
(No Transcript)
40
The Canny Edge Detector
41
The Canny Edge Detector

• The Canny ED is very useful for finding contours.
• It takes as an input s which is used to apply a
  Gaussian filter on the image as the first step.
• Gradient magnitude and direction are found on the
  smoothed image.
• Gradient direction is used to thin edges by
  suppressing any pixel response that is not higher
  than the two neighboring pixels on either side of
  the direction. This is called non-maximal
  suppression.

42

• The 2 8-neighbors of a pixel that are to be
  compared are found by rounding off the computed
  gradient to yield one pixel on either side of the
  center pixel.
• Once the gradient magnitudes are thinned. High
  magnitude contours are tracked.
• In the final stage continuous contour segments
  are sequentially followed.
• Contour following is initiated only on edge
  pixels where gradient magnitude meets a high
  threshold however once started a contour may be
  followed through pixels whose gradient magnitude
  are lower than the threshold.

43
(No Transcript)
44
(No Transcript)
45
(No Transcript)
46

• Segmentation Part II