Subject Name Digital Image Processing
Code : MTCS-202 (C)
[Link] (Computer Science )
2nd Semester
Faculty Name: Dr Satya Bhushan Verma
Image Segmentation
Image segmentation is the process of partitioning an image into multiple segments. Image
segmentation is typically used to locate objects and boundaries in images.
Segmentation partitions an image into distinct regions containing each pixels with similar
attributes. To be meaningful and useful for image analysis and interpretation, the regions
should strongly relate to depicted objects or features of interest. Meaningful segmentation is
the first step from low-level image processing transforming a greyscale or colour image into
one or more other images to high-level image description in terms of features, objects, and
scenes. The success of image analysis depends on reliability of segmentation, but an accurate
partitioning of an image is generally a very challenging problem.
Why do we need Image Segmentation?
Image segmentation breaks the image down into meaningful regions. For example, in the
picture below you can clearly see segments corresponding to the sky, the trees, the elephant
and the grass.
Figure: Segmented image
Image segmentation is a vital part of image analysis process. It differentiates between the
objects we want to inspect further and the other objects or their background.
Segmentation
Contextual segmentation or Non-Contextual
Region growing Thresholding
Pixel connectivity Simple Thresholding
growing Split-and-merge Adaptive thresholding
Region similarity Adaptive Thresholding
segmentation Colour thresholding
Region growing Colour Thresholding
Split-and-merge Segmentation
Figure: Classification of Segmentation
Segmentation techniques are either contextual or non-contextual
1. Contextual segmentation or Region growing: Non-contextual thresholding
groups pixels with no account of their relative locations in the image plane.
Contextual segmentation can be more successful in separating individual objects
because it accounts for closeness of pixels that belong to an individual object. Two
basic approaches to contextual segmentation are based on signal discontinuity or
similarity. Discontinuity-based techniques attempt to find complete boundaries
enclosing relatively uniform regions assuming abrupt signal changes across each
boundary. Similarity-based techniques attempt to directly create these uniform regions
by grouping together connected pixels that satisfy certain similarity criteria. Both the
approaches mirror each other, in the sense that a complete boundary splits one region
into two.
1.1. Pixel connectivity: Pixel connectivity is defined in terms of pixel neighbourhoods.
A normal rectangular sampling pattern producing a finite arithmetic lattice {(x,y): x =
0, 1, ..., X−1; y = 0, 1, ..., Y−1} supporting digital images allows us to define two
types of neighbourhood surrounding a pixel. A 4-neighbourhood {(x−1,y), (x,y+1),
(x+1,y), (x,y−1)} contains only the pixels above, below, to the left and to the right of
the central pixel (x,y). An 8-neighbourhood adds to the 4-neighbourhood four
diagonal neighbours: {(x−1,y−1),(x−1,y), (x−1,y+1), (x,y+1), (x+1,y+1), (x+1,y),
(x+1,y−1), (x,y−1)}
Figure : 4- Neighbourhood and 8-Neighbourhood
One of the simplest and most common algorithms for labelling connected regions after
greyscale or colour thresholding exploits the "grassfire" or "wave propagation" principle:
after a "fire" or "wave" starts at one pixel, it propagates to any of the pixel's 4- or 8-
neighbours detected by thresholding. Each already visited (i.e. "burnt away" or "wet") pixel
cannot be visited again, and after the entire connected region is labelled, its pixels are
assigned a region number, and the procedure continues to search for the next connected
region. Magenta and yellow stars below indicate the fire, or wave front and the burnt away
pixels, respectively. To label a region, the fire starts from its first chosen pixel:
Figure: 4- and 8-connectivity produce different segmentation results:
Figure: 4-Connected pixel
Figure: 8-Connected pixel
1.2. Region similarity: The uniformity or non-uniformity of pixels to form a connected
region is represented by a uniformity predicate, i.e. a logical statement, or condition
being true if pixels in the regions are similar with respect to some property (colour,
grey level, edge strength, etc). A common predicate restricts signal variations over a
neighbourhood: the predicate P(R), where R denotes a connected region, is TRUE if
|f(x,y) − f(x+ξ,y+η)| ≤ Δ and FALSE otherwise (here, (x,y) and (x+ξ,y+η) are the
coordinates of neighbouring pixels in region R. This predicate does not restrict the
grey level variation within a region because small changes in signal values can
accumulate over the region.
Intra-region signal variations can be restricted with a similar predicate: P(R) = TRUE
if |f(x,y) − &muR| ≤ &Delta and FALSE otherwise where (x,y) is a pixel from the
region R and μR is the mean value of signals f(x,y) over the entire region R.
1.3. Region growing: The bottom-up region growing algorithm starts from a set of seed
pixels defined by the user and sequentially adds a pixel to a region provided that the
pixel has not been assigned to any other region, is a neighbour of that region, and its
addition preserves uniformity of the growing region.
This type of segmentation is simple but unstable. It is very sensitive to a chosen
uniformity predicate, i.e. small changes of the uniformity threshold may result in large
changes of the regions found. Also, very different segmentation maps are obtained
under different routes of scanning an image, different modes of exhausting neighbours
of each region, different seeds, and different types of pixel connectivity.
Generally, a "good" complete segmentation must satisfy the following criteria:
1. All pixels have to be assigned to regions.
2. Each pixel has to belong to a single region only.
3. Each region is a connected set of pixels.
4. Each region has to be uniform with respect to a given predicate.
5. Any merged pair of adjacent regions has to be non-uniform.
1.4. Split-and-merge segmentation: The top-down split-and-merge algorithm considers
initially the entire image to be a single region and then iteratively splits each region
into subregions or merges adjacent regions until all regions become uniform or until
the desired number of regions have been established. A common splitting strategy for
a square image is to divide it recursively into smaller and smaller quadrants until, for
any region R, the uniformity predicate P(R) is TRUE. The strategy builds a top-down
quadtree: if P(image) is FALSE, the image is divided into four quadrants; if
P(quadrant) is FALSE, the quadrant is divided into subquadrants; and so on:
The splitting stage alternates with a merging stage, in which two adjacent
regions Ri and Rj are combined into a new, larger region if the uniformity predicate
for the union of these two regions, P(Ri ∪ Rj), is TRUE.
2. Non-contextual thresholding: Thresholding is the simplest non-contextual
segmentation technique. With a single threshold, it transforms a greyscale or colour
image into a binary image considered as a binary region map. The binary map
contains two possibly disjoint regions, one of them containing pixels with input data
values smaller than a threshold and another relating to the input values that are at or
above the threshold. The former and latter regions are usually labelled with zero (0)
and non-zero (1) labels, respectively. The segmentation depends on image property
being thresholded and on how the threshold is chosen.
Generally, the non-contextual thresholding may involve two or more thresholds as
well as produce more than two types of regions such that ranges of input image
signals related to each region type are separated with thresholds. The question of
thresholding is how to automatically determine the threshold value.
2.1. Simple thresholding: The most common image property to threshold is pixel grey
level: g(x,y) = 0 if f(x,y) < T and g(x,y) = 1 if f(x,y) ≥ T, where T is the threshold.
Using two thresholds, T1 < T1, a range of grey levels related to region 1 can be
defined: g(x,y) = 0 if f(x,y) < T1 OR f(x,y) > T2 and g(x,y) = 1 if T1 ≤ f(x,y) ≤ T2.
2.2. Adaptive thresholding: Threshold separates the background from the object, the
adaptive separation may take account of empirical probability distributions of object
(e.g. dark) and background (bright) pixels. Such a threshold has to equalise two kinds
of expected errors: of assigning a background pixel to the object and of assigning an
object pixel to the background. More complex adaptive thresholding techniques use a
spatially varying threshold to compensate for local spatial context effects (such a
spatially varying threshold can be thought as a background normalisation).
A simple iterative adaptation of the threshold is based on successive refinement of the
estimated peak positions. It assumes that (i) each peak coincides with the mean grey
level for all pixels that relate to that peak and (ii) the pixel probability decreases
monotonically on the absolute difference between the pixel and peak values both for
an object and background peak. The classification of the object and background
pixels is done at each iteration j by using the threshold Tj found at previous iteration.
Thus, at iteration j, each grey level f(x,y) is assigned first to the object or background
class (region) if f(x,y) ≤ Tj or f(x,y) > Tj, respectively. Then, the new
threshold, Tj+1 = 0.5(μj,ob + μj,bg) where μj,ob and μj,bg denote the mean grey level
at iteration j for the found object and background pixels, respectively:
2.3. Colour thresholding: Color segmentation may be more accurate because of more
information at the pixel level comparing to greyscale images. The standard Red-
Green-Blue (RGB) colour representation has strongly interrelated colour components,
and a number of other colour systems (e.g. HSI Hue-Saturation-Intensity) have been
designed in order to exclude redundancy, determine actual object / background
colours irrespectively of illumination, and obtain more more stable segmentation. An
example below shows that colour thresholding can focus on an object of interest
much better than its greyscale analogue:
Segmentation of colour images involve a partitioning of the colour space, i.e. RGB or
HSI space. One simple approach is based on some reference (or dominant) colour
(R0, G0, B0) and thresholding of Cartesian distances to it from every pixel
colour f(x,y) = (R(x,y),G(x,y),B(x,y)):
where g(x,y) is the binary region map after thresholding. This thresholding rule
defines a sphere in RGB space, centred on the reference colour. All pixels inside or on
the sphere belong to the region indexed with 1 and all other pixels are in the region 0.
Also, there can be an ellipsoidal decision surface if independent distance thresholds
are specified for the R, G, and B components. Generally, colour segmentation, just as
the greyscale one, may be based on the analysis of 3D colour histograms or their more
convenient 2D projections. A colour histogram is built by partitioning of the colour
space onto a fixed number of bins such that the colours within each bin are considered
as the same colour. An example below of the partitioned 11×11×11 RGB colour space
is from
Figure: 11×11×11 RGB colour space
How fine should be the partitioning depends on the application domain. In many cases
colour segmentation exploits only a few dominant colours corresponding to distinct
peaks of the pixel-wise colour distribution
Figure : Colour image “baboon”, and its colour 6×6×6 histogram
If a chosen colour space separates colourless intensity values from intensity-
independent colour components (such as hue and saturation or normalised red / blue
colurs), colour segmentation can be based on a few pre-selected colours, e.g. on the
eight primary colours (black, red, green, blue, yellow, cyan, magenta, white). An
example below shows a digitised picture of the Rembrandt's canvas "Doctor Nicolaes
Tulp's Demonstration of the Anatomy of the Arm" its 8-bin histogram of the primary
colours, and the corresponding colour regions:
References:
[1]. Image Segmentation,
[Link]
html/[Link]
[2]. Image Segmentation, [Link]
34430371
[3]. Colour Analysis, [Link]