IMAGE SEGMENTATION
TYPES OF SEGMENTATION
1. THRESHOLD BASED SEGMENTATION
The bottom left and right
images show examples of
oversegmentation and
undersegmentation
respectively
Threshold value 0
Threshold value 10
Threshold value 25
Threshold value 50
Threshold value 75
Threshold value 100
Threshold value 125
Threshold value 150
Threshold value 175
Threshold value 200
Threshold value 215
Threshold value 230
HISTOGRAM FOR IMAGES
Plain Image Noise Image
Threshold t may be
midway between two peaks with grey values p1 and p2
HISTOGRAM OVERLAP
Minimum choice for t is usually better than the
midway choice.
We may select a threshold that maximizes the
variance between objects and background, or one
that attempts to optimize both these “within” and
“between” variances.
MINIMUM VARIANCE
For simplicity, we normalize the histogram H(v) of the
image to a proper distribution h(v) as
h(v) = 1.
This can be achieved by setting
h(v) = H(v)/n for all v,
where n is the total number of pixels in the image.
The variance of the grey values σ2 in the image then
(by definition) equals
ITERATIVE THRESHOLDING
Problem: noisy histograms
(salt-and pepper noise)
OPTIMAL THRESHOLDING
Optimal thresholding is a technique that approximates the
histogram using a weighted sum of distribution functions, and
then sets a threshold in such a way that the number of
incorrectly segmented pixels (as predicted from the
approximation) is minimal.
If the histogram is bimodal (two segments; objects and
background), and that the histogram can be
approximated well using a sum of two Gaussians. In
this case, this means we model the normalized
histogram h as
where g0 and g1 are Gaussian functions with unknown mean
and variance, and P0 and P1 are the global probabilities also
unknown that a pixel belongs to segment 0 or 1 respectively.
Note that P0 + P1 = 1.
ENHANCING THRESHOLD SEGMENTATION
Noisy Image Threshold Segmentation Histogram of
original image
Filtering should be applied before or after
the thresholding
Median filter Applied Thresholding Result
PRE-OR-POST PROCESSING
Noisy Original Thresholded Image
Thresholding Fails in this case
Image Closing
Thresholded Result Removing some small
using 7*7 square
area segments
Image Closing
using 9*9 square
NEED FOR LOCAL THRESHOLDING
Global threshold: The threshold was the same for the
entire image.
Pixel Pixel
value 240 value 120
STEPS:
If we divide the image into n subimages.
Then the grey value of bone pixels in each subimage
will be relatively constant, because the gradient
effect is comparatively.
For each subimage, we can now find an appropriate
threshold to extract the bone pixels.
Merging the segmentation results of each subimage
into a complete image again will give us a good
segmentation of the whole image.
SUMMARY
Advantages:
Simple to implement
Fast (especially if repeating on similar images)
Good for some kinds of images (e.g.,documents,
controlled lighting)
Disadvantages:
No guarantees of object coherency may have holes,
extraneous pixels, etc. (incomplete) solution: post-
processing with morphological operators
EDGE BASED SEGMENTATION
There are three basic types of gray-level
discontinuities in a digital image: points, lines, and
edges
The most common way to look for discontinuities is to
run a mask through the image.
We say that a point, line, and edge has been detected
at the location on which the mask is centered
if ,where
Point detection
a point detection mask
Line detection
a line detection mask
Edge detection: Gradient
operation
f
f fx
Gx
Gy
y
1
f mag (f ) Gx Gy 2 2 2
Gy
( x, y ) tan (
1
)
Gx
Edge detection:
Laplacian operation
2 f 2 f
f 2 2
2
x y
r2
r 2 2 2 2
h(r )
2
e
4
REGION-BASED SEGMENTATION
SRG
Region growing: Groups pixels or sub-region into
larger regions.
step1:
Start with a set of “seed” points and from
these grow regions by appending to each
seed those neighboring pixels that have
properties similar to the seed.
step2:
Region splitting and merging
4
1
REGION-BASED SEGMENTATION
SRG
Advantage:
With good connectivity
Disadvantage:
Initial seed-points:
different sets of initial seed-point cause different segmented
result
Time-consuming problem
4
2
REGION-BASED SEGMENTATION
USRG
Unseeded region growing:
no explicit seed selection is necessary, the
seeds can be generated by the
segmentation procedure automatically.
It is similar to SRG except the choice of
seed point
4
3
REGION-BASED SEGMENTATION
USRG
Advantage:
easy to use
can readily incorporate high level knowledge of the
image composition through region threshold
Disadvantage:
slow speed
4
4
REGION-BASED SEGMENTATION
FAST SCANNING(1)
Fast scanning Algorithm:
The fast scanning
algorithm somewhat
resembles unseeded
region growing
the number of clusters of
both two algorithm would
not be decided before
image passing through
them.
45
REGION-BASED SEGMENTATION
FAST SCANNING(2)
4
6
REGION-BASED SEGMENTATION
FAST SCANNING(3)
Last step:
merge small region to
big region
47
REGION-BASED SEGMENTATION
FAST SCANNING
Advantage:
The speed is very fast
The result of segmentation will be intact with good
connectivity
Disadvantage:
The matching of physical object is not good
It can be improved by morphology and geometric
mathematic
4
8
REGION-BASED SEGMENTATION
FAST SCANNING-IMPROVED BY MORPHOLOGY
erosion
dilation
A B {c E N | c a b for some a A and b B} A ! B {x E N x b A for every b B}
49
REGION-BASED SEGMENTATION
FAST SCANNING-IMPROVED BY
MORPHOLOGY
dilation erosion
50
REGION-BASED SEGMENTATION
FAST SCANNING-IMPROVED BY MORPHOLOGY
opening closing
51
Erosion=>Dilation Dilation=>Erosion
CLUSTERING SEGMENTATION
WHAT IS CLUSTER ANALYSIS?
Finding groups of objects such that the objects in a group
will be similar (or related) to one another and different from
(or unrelated to) the objects in other groups
Inter-cluster
Intra-cluster distances are
distances are maximized
minimized
NOTION OF A CLUSTER CAN BE AMBIGUOUS
How many clusters? Six Clusters
Two Clusters Four Clusters
TYPES OF CLUSTERS: CONTIGUITY-BASED
Contiguous Cluster (Nearest neighbor or
Transitive)
A cluster is a set of points such that a point in a cluster is
closer (or more similar) to one or more other points in the
cluster than to any point not in the cluster.
8 contiguous clusters
TYPES OF CLUSTERS: DENSITY-BASED
Density-based
A cluster is a dense region of points, which is separated by
low-density regions, from other regions of high density.
Used when the clusters are irregular or intertwined, and
when noise and outliers are present.
6 density-based clusters
EUCLIDEAN DENSITY – CENTER-BASED
Euclidean density is the number of points within
a specified radius of the point
K-MEANS
Choose a fixed number of Algorithm
clusters fix cluster centers; allocate
points to closest cluster
fix allocation; compute
Choose cluster centers and
best cluster centers
point-cluster allocations to
minimize error x could be any set of
features for which we can
can’t do this by search,
compute a distance
because there are too
(careful about scaling)
many possible allocations.
2
iclusters jelements of i'th cluster
x j i
Image Clusters on intensity Clusters on color
K-means clustering using intensity alone and color alone
Image Clusters on color
K-means using color alone, 11 segments
K-means using
color alone,
11 segments.
LIMITATIONS OF K-MEANS
K-means has problems when clusters are
of differing
Sizes
Densities
Non-globular shapes
K-means has problems when the data
contains outliers.
LIMITATIONS OF K-MEANS: DIFFERING
SIZES
Original Points K-means (3 Clusters)
LIMITATIONS OF K-MEANS: DIFFERING
DENSITY
Original Points K-means (3 Clusters)
LIMITATIONS OF K-MEANS: NON-GLOBULAR
SHAPES
Original Points K-means (2 Clusters)
OVERCOMING K-MEANS LIMITATIONS
Original Points K-means Clusters
One solution is to use many clusters.
Find parts of clusters, but need to put together.
OVERCOMING K-MEANS LIMITATIONS
K-means Clusters
MATCHING
If we want to locate an object in an image, and we
have available an example of what it should look like
(a template), we can find this object by matching the
template to various image locations until we have
found the object.
The most straightforward way of determining whether
a template ‘fits’ would be to place the template at a
certain image location, and see whether the grey
values of the template and the underlying image
image grey values all match.
However, because there will generally be some
differences between the image and template values
because of noise and other artifacts, this is not a very
practical method. More useful is a quantitative measure
of fit such as
where f is the image, g the M × N template, and the
variables p and q determine the location of the template
in the image. This measure will be small if the template
is similar to the part of the image under investigation;
then all grey value differences g(x, y)−f(x+p, y +q) are
small and the sum M1 will be small. The location of
optimal template fit is found by minimizing M1 to p and
q.
Image matching. The left image gives a template of the object we wish to
locate in the right image. This is achieved by matching of the template to
various locations in the right image.
