Image segmentation
1
Context
{ Segmentation decomposes the image into parts for
further analysis
z Example: background subtraction in human
motion analysis
{ Once the region of interest is segmented, the
representation space can be changed (from image-
space to feature space)
Circumscribed (benign) Spiculated (malignant)
lesions in digital lesions in digital
2
mammography mammography
What is segmentation ?
{ Partitioning an image into regions
corresponding to objects
{ All pixels in a region share a
common property
{ Simplest property that pixels can
share: intensity
{ Thresholding=separation of light
and dark regions
3
A classification of segmentation
techniques
{ Intensity-based segmentation:
Thresholding
{ Edge-based segmentation
{ Region-based segmentation
4
A classification of segmentation
techniques
{ Intensity-based segmentation:
Thresholding
{ Edge-based segmentation
{ Region-based segmentation
5
Assumptions for thresholding
1. the intensity values are different in different regions
2. within each region, which represents the
corresponding object in a scene, the intensity values
are similar.
6
Intensity-based thresholding
Image thresholding classifies pixels into two categories:
– Those to which some property measured from the image
falls below a threshold, and those at which the property
equals or exceeds a threshold.
– Thresholding creates a binary image : binarization
e.g. perform cell counts in histological images
Choosing a threshold is a critical task.
n=imread(‘nodules1.jpg’);
figure(1); imshow(n);
n1=im2bw(n,0.35);
n2=im2bw(n,0.75);
figure(2), imshow(n1);
figure(3), imshow(n2); 7
Fixed versus dynamic thresholding
• In fixed (or global) thresholding, the threshold value is
held constant throughout the image:
0 f(x,y)<T
g(x,y) = 1 f(x,y)>=T
{ Local (or dynamic thresholding): depends on the
position in the image. The image is divided into
overlapping sections which are thresholded one by one.
8
Threshold detection methods
{ P-tile thresholding
{ Optimal thresholding
{ Mixture modelling
{ Adaptive thresholding
{ Note: all the above methods are
automated
9
P-tile method
{ a priori information: object is brighter/darker than
background and occupies a certain known percentile
1/p from the total image area (example: printed text
sheet)
{ We set the threshold by finding the intensity level
such that 1/p image pixels are below this value
{ We use the cumulative histogram
g
c (g ) = ∑ h( k )
k =0
n
h( k ) = k
n
{ T verifies the equation c(T)=1/p (for a dark
foreground)
{ c(T)=1-1/p (for a bright foreground)
10
Finding modes
{ Histogram shape analysis
{ Foreground pixels form one peak
{ Background pixels form the second peak
{ Intuitively: the threshold is set as the gray level that
has a minimum value between two maxima
{ Problem: noisy histograms (salt-and pepper noise)
11
Optimal thresholding
{ Idea: the histogram of an image is
the sum of two overlapping
distributions
{ Optimal threshold: overlapping
point of these distributions
(corresponds to the minimum
probability between the maxima of
2 distributions)
{ Problem: distributions are unknown
12
Comparison between conventional and
optimal thresholding
13
Optimal thresholding by clustering
{ Simplest case: segmentation into two
classes (object/background).
{ The intensities in each class will be our
clusters.
{ We want to find a threshold such that
each pixel on each side of the threshold is
closer in intensity to the mean of all
pixels on that side of the threshold than
to the mean of all pixels on the other side
of the threshold.
14
Iterative optimal threshold selection
1. Select an initial estimate for T
2. Segment the image using T. This
produces 2 groups: G1 pixels with value >T
and G2 , with value <T
3. Compute µ1 and µ2, average pixel values
of G1 and G2
4. New threshold: T=1/2(µ1+µ2)
5. Repeat steps 2 to 4 until T stabilizes.
15
Iterative Clustering Algorithm
m1(1)=260.83, m2(1)=539.00
m1(2)=39.37, m2(2)=1045.65
m1(3)=52.29, m2(3)=1098.63
m1(4)=54.71, m2(4)=1106.28
m1(5)=55.04, m2(5)=1107.24
m1(6)=55.10, m2(6)=1107.44
m1(7)=55.10, m2(7)=1107.44
16
Optimal thresholding : the Otsu method
{ Optimal thresholding methods select
the threshold based on the
minimization of a criterion function.
{ The criterion for Otsu is the
minimization of the within-group
variance of the two groups of pixels
separated by the threshold.
17
18
Mixture modelling
{ Assumption: region intensities are each
normal distributions (Gaussians)
19
Mixture modelling (cont’d)
{ Each of the Gaussian distributions has a mean and
standard deviation independent of the threshold that we
choose
{ Foreground/background case:
{ We need to estimate 6 parameters
{ Evaluation of how well the sum of the distributions
approximate the histogram
{ The parameters will be chosen such as to minimize the
error F
20
Thresholding and illumination
21
Connected component labeling
From Brian Morse, http://morse.cs.byu.edu/650/home/index.php
22
Connected component labeling
{ 1. Scan through the image pixel by pixel across
each row in order:
z If the pixel has no connected neighbors with the
same value that have already been labeled, create
a new unique label and assign it to that pixel.
z If the pixel has exactly one label among its
connected neighbor with the same value that has
already been labeled, give it that label.
z If the pixel has two or more connected neighbors
with the same value but different labels, choose one
of the labels and remember that these labels are
equivalent.
{ 2. Resolve the equivalencies by making another
pass through the image and labeling each pixel
with a unique label for its equivalence class.
From Brian Morse, http://morse.cs.byu.edu/650/home/index.php
23
Thresholding: 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
24
Next lecture
• Edge-based segmentation
• Region-based segmentation
25
