Image feature extraction

IMAGE REPRESENTATION
&
FEATURE EXTRACTION
ATSIP
6/3/2014 ]2ushin $hah 1

Representation
 Representation means that we make the object information more accessible
for computer-interpretation .
 Two types of representation
 Using boundary (External characteristics)
 Using pixels of region (Internal characteristics)
6/3/2014 ]2ushin $hah 3

Description
 Description means that we quantify our representation of the object
 Boundary Descriptors
 Geometrical descriptors : Diameter, perimeter, eccentricity, curvature
 Shape Numbers
 Fourier Descriptors
 Statistical Moments
 Regional Descriptors
 Geometrical descriptors: Area, compactness, Euler number
 Texture
 Moments of 2D Functions
6/3/2014 ]2ushin $hah 4

Desirable properties of descriptors
 They should define a complete set
 Two objects must have the same descriptors if and only if they have the same shape .
 They should be invariant to Rotation, Scaling and Translation (RST)
 They Should be a compact set
 A descriptor should only contain information about what makes an object unique,
or different from the other objects.
 The quantity of information used to describe this characterization should be less
than the information necessary to have a complete description of the object itself.
 They should be robust
 Work well against Noise and Distortion
 They should have low computational complexity
6/3/2014 ]2ushin $hah 5

Introduction
 The common goal of feature extraction and representation techniques is to
convert the segmented objects into representations that better describe their main
features and attributes.
 The type and complexity of the resulting representation depend on many
factors, such as
 the type of image (e.g., binary, gray-scale, or color),
 the level of granularity (entire image or individual regions) desired, and
 the context of the application that uses the results (e.g., a two-class
pattern classifier that tells circular objects from noncircular ones or an
image retrieval system that retrieves images judged to be similar to an
example image).
6/3/2014 ]2ushin $hah 6

Introduction
 “Feature extraction is the process by which certain features of interest
within an image are detected and represented for further processing.”
 It is a critical step in most computer vision and image processing solutions
because it marks the transition from pictorial to non-pictorial
(alphanumerical, usually quantitative) data representation.
 The resulting representation can be subsequently used as an input to a
number of pattern recognition and classification techniques, which will
then label, classify, or recognize the semantic contents of the image or its
objects.
6/3/2014 ]2ushin $hah 7

FEATURE VECTORS & VECTOR SPACES
 A feature vector is a n × 1 array that encodes the n features (or
measurements) of an image or object.
 The array contents may be
 symbolic (e.g., a string containing the name of the predominant color in
the image),
 numerical (e.g., an integer expressing the area of an object, in pixels),
 or both.
 Mathematically, a numerical feature vector x is given by
𝑥 = 𝑥1, 𝑥2, … , 𝑥 𝑛
𝑇
 where n is the total number of features and
 T indicates the transpose operation.
6/3/2014 ]2ushin $hah 8

FEATURE VECTORS & VECTOR SPACES
 The feature vector is a compact
representation of an image (or object
within the image), which can be
associated with
 the notion of a feature space,
 an n-dimensional hyperspace that
allows the visualization (for n <
4) and
 interpretation of the feature
vectors’ contents, their relative
distances, and so on.
6/3/2014 ]2ushin $hah 9

Invariance & Robustness
 A common requirement for feature extraction and representation techniques
is that the features used to represent an image be invariant to rotation,
scaling, and translation, collectively known as RST.
 RST invariance ensures that a machine vision system will still be able to
recognize objects even when they appear at different size, position within the
image, and angle (relative to a horizontal reference).
6/3/2014 ]2ushin $hah 10

Binary Object Features
 A binary object is a connected region within a binary image f (x, y), which
will be denoted as 𝑂𝑖, i > 0.
 Mathematically, we can define a function 𝑂𝑖(𝑥, 𝑦) as follows:
 𝑂𝑖 𝑥, 𝑦 =
1 𝑖𝑓(𝑓(𝑥, 𝑦) ∈ 𝑂𝑖)
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
Area
6/3/2014 ]2ushin $hah 11

BOUNDARYDESCRIPTORS
 In this section, we will look at contour-based representation and description
techniques.
 These techniques assume that the contour (or boundary) of an object can be
represented in a convenient coordinate system (Cartesian—the most
common, polar, or tangential) and rely exclusively on boundary pixels to
describe the region or object.
 Object boundaries can be represented by different techniques, ranging from
simple polygonal approximation methods to more elaborated techniques
involving piecewise polynomial interpolations such as B-spline curves.
6/3/2014 ]2ushin $hah 12

BOUNDARYDESCRIPTORS
 The techniques described in this section assume that the pixels belonging to
the boundary of the object (or region) can be traced, starting from any
background pixel, using an algorithm known as bug tracing that works as
follows:
 As soon as the conceptual bug crosses into a boundary pixel, it makes a
left turn and moves to the next pixel; if that pixel is a boundary pixel, the
bug makes another left turn, otherwise it turns right; the process is
repeated until the bug is back to the starting point.
 As the conceptual bug follows the contour, it builds a list of coordinates of
the boundary pixels being visited.
6/3/2014 ]2ushin $hah 13

Chain Code, Freeman Code, & Shape Number
 Chain codes are alternative methods for tracing and describing a contour.
 A chain code is a boundary representation technique by which “A contour is
represented as a sequence of straight line segments of specified length
(usually 1) and direction”.
 The simplest chain code mechanism, also known as crack code, consists of
assigning a number to the direction followed by a bug tracking algorithm as
follows: right (0), down (1), left (2), and up (3).
 By allocating numbers based on directions, the boundary of an object is
reduced to a sequence of numbers .
6/3/2014 ]2ushin $hah 14

Chain Code
 Steps for construction chain codes
 Select some starting point of the boundary and represent it by its absolute
coordinates in the image
 Represent every consecutive point by a chain code showing transition
needed to go from current point to next point on the boundary
 Stop if the next point is the initial point or the end of the boundary
6/3/2014 ]2ushin $hah 15

 Assuming that the total
number of boundary points is p
(the perimeter of the contour),
the array C (of size p), where
C(p) = 0, 1, 2, 3, contains the
chain code of the boundary.
 A modified version of the basic
chain code, known as the
Freeman code, uses eight
directions instead of four.
 Figure 18.10 shows an example
of a contour, its chain code, and
its Freeman code.
6/3/2014 ]2ushin $hah 16

 Once the chain code for a boundary has been computed, it is possible to
convert the resulting array into a Rotation-Invariant Equivalent, known as the
first difference.
 It is obtained by encoding the number of direction changes, expressed in
multiples of 90◦ (according to a predefined convention, for example, counter
clockwise), between two consecutive elements of the Freeman code.
 The first difference of Smallest magnitude is obtained by treating the
resulting array as a circular array and rotating it cyclically until the resulting
numerical pattern results in the smallest possible number is known as the
Shape number of the contour.
6/3/2014 ]2ushin $hah 17

 The shape number is Rotation invariant and Insensitive to the starting point
used to compute the original sequence.
 Figure 18.11 shows an example of a contour, its chain code, first differences,
and shape number.
6/3/2014 ]2ushin $hah 18

Shape Number
The shape number of a boundary is defined as the first difference of smallest
magnitude
The order n of a shape number is defined as the number of digits in its
representation
6/3/2014 ]2ushin $hah 19

Algorithm formaking a shape number
 Goal: To represent a given boundary by a shape number of order n
 Step-1: Obtain the major axis of the shape and consider it as one of the
coordinate axis
 Step-2: Find the basic (smallest) rectangle that has sides parallel to major
axis and just covers the shape
 Step-3: From possible rectangles of order n, find one which best
approximates rectangle of step-2
 Step-4: Orient the rectangle, so that its major axis coincides with that of
the shape
 Step-5: Obtain the first difference chain code of minimum magnitude after
circular shift
6/3/2014 ]2ushin $hah 20

Chain Code
 Advantages
 Preserves the information of interest
 Provides good compression of boundary description
 They are translation invariant
 Problems
 Long chains of codes
 No invariance to Rotation and Scale
 Sensitive to Noise
 Solution
 Re-sample the image to a lower resolution before calculating the code
6/3/2014 ]2ushin $hah 22

Chain Code
 Problem
 A chain code sequence depends on a starting point.
 Solution
 Treat a chain code as a circular sequence and redefine the starting point so
that the resulting sequence of numbers forms an integer of minimum
magnitude after circular shift
2 2 3 0  0 2 2 3
 The first difference of a chain code is counting the number of direction
change (in counter clockwise) between 2 adjacent elements of the code
6/3/2014 ]2ushin $hah 24

PolygonalApproximation
 “Approximates the boundary by a set of connected line segments “
 Polygonal approximation provides a simple representation of the
Planar Object Boundary.
 Mathematical Definition
 Let the set of points of boundary be 𝑿 = 𝒙 𝟏, 𝒙 𝟐, … , 𝒙 𝒏
 Divide this set into segment Ʌ = 𝝀 𝟏, 𝝀 𝟐, … , 𝝀 𝒏
 Approximate each segment by straight line by minimization of objective
function 𝑱 = 𝒊 𝒅 𝒙𝒊, 𝒍𝒋 , 𝒙𝒊 ∈ 𝝀𝒋
6/3/2014 ]2ushin $hah 26

PolygonalApproximation
 Approximation leads to Loss of Information
 The number of straight line segments used determines the accuracy of the
approximation
 For a closed boundary, approximation becomes exact when no. of
segments of the polygons is equal to the no. of points in the boundary
 However, the goal of approximation is
 To capture the essence of the object shape with minimal loss
 Thus, it saves the no. of bytes required for boundary representation
6/3/2014 ]2ushin $hah 27

The Split Method (Top-down)
 Iteratively decompose a boundary into a set of small segments and
represent the segment by a straight line
 Algorithm:
 Step-1: Take the line segment connecting the end points of the boundary
(if the boundary is closed, consider the line segment connecting the two
farthest points).
 Step-2: Find the boundary point with maximum distance from the line
segment
 Step-3: If the distance is above threshold, split the segment into two
segments at that point (i.e., new vertex).
 Step-4: Repeat the same procedure for each of the two sub segments until
the distance is below threshold
6/3/2014 ]2ushin $hah 28

The Merge Method (Bottom-up)
 Operate in a direction opposite to that of splitting method
 Algorithm:
 Step-1: Use the first two boundary points to define a line segment
 Step-2: Add a new point if it does not deviate too far from the current line
segment
 Step-3: Update the parameters of the line segment using least-squares
 Step-4: Start a newline segment when boundary points deviate too far
from the line segment
6/3/2014 ]2ushin $hah 30

The Split & MergeAlgorithm
 Problems of the split and merge methods
 Depending on threshold , vertices of polygon not necessarily correspond
to points of inflections (such as corners) in the boundary
 Combine split and merge method
 After recursive subdivision (split), allow adjacent segments to be replaced
by a single segment (merge)
6/3/2014 ]2ushin $hah 31

Signatures
 “Signature is a 1D representation of a boundary”
 It is obtained by representing the boundary in a polar coordinate system then
 Computing the distance r between each Pixel along the boundary and the
Centroid of the region, and
 The angle θ subtended between a straight line connecting the boundary
pixel to the centroid and a horizontal reference (Figure 18.12, top).
 The resulting plot of all computed values for 0 ≤ θ ≤ 2π (Figure 18.12,
bottom) provides a concise representation of the boundary that is translation
invariant can be made rotation invariant (if the same starting point is always
selected), but is not scaling invariant.
 Figure 18.13 illustrates the effects of noise on the signature of a contour.
6/3/2014 ]2ushin $hah 32

Signatures
6/3/2014 ]2ushin $hah 33

Signature
 Signatures are invariant to location, but will depend on rotation and
scaling.
 Rotation invariance can be improved by selecting a unique
starting point (e.g. based on major axis)
 Scale invariance can be achieved by normalizing amplitude of
signature (divide by variance)
6/3/2014 ]2ushin $hah 34

Skeletons
 Skeletons produce a one pixel wide graph that has the same basic shape of
the region, like a stick figure of a human
 Hence, they provide a compact and often highly intuitive representation
 It can be used to analyse the geometric structure of a region
 Also popular tool in object recognition
6/3/2014 ]2ushin $hah 35

MedialAxis Transform (MAT)
 Provides skeleton of an object.
 The MAT of a region R with border B is defined as follows:
 For each point p of R, we find its closest neighbour in B.
 If p has more then one such points, it is said to belong to the medial axis
(skeleton) of R .
6/3/2014 ]2ushin $hah 36

6/3/2014 ]2ushin $hah 37

 Medial Axis augmented by radius function & Transformation is invertible
 The medial axis of a circle is its center.
 the medial axis of an ellipse is its center (the midpoint of the line that
connects the two foci of the ellipse), too.
 Equilateral triangle : the segments connecting the middle of the bases and
the center of the figure.
 Arbitrary triangle : the segments connecting the middle of the sides with the
center of gravity (where all medians cross).
6/3/2014 ]2ushin $hah 38

 Application
 Shape matching
 Animation
 Dimension reduction
 Solid modelling
 Smoothing or sharpening of shape
 Motion planning
 Mesh generation
6/3/2014 ]2ushin $hah 39

Boundary Descriptors
 There are several simple geometric measures that can be useful for
describing a boundary
 Length
 the number of pixels along a boundary gives a rough approximation of its
length
 For a chain-coded curve with unit spacing
 Length = the number of vertical and horizontal components + √2 * the
number of diagonal components
 Diameter (Major Axis)
 𝐷𝑖𝑎𝑚 𝐵 = max
𝑖,𝑗
[𝐷(𝑝𝑖, 𝑝𝑗)]
6/3/2014 ]2ushin $hah 40

Boundary Descriptors
 Minor Axis
 the line perpendicular to the major axis
 Eccentricity
 Ratio of major axis to minor axis
6/3/2014 ]2ushin $hah 41

Fourier Descriptors
 The idea behind Fourier descriptors is to traverse the pixels belonging to a
boundary, starting from an arbitrary point, and record their coordinates.
 Each value in the resulting list of coordinate pairs
𝑥0, 𝑦0 , 𝑥1, 𝑦1 , … , 𝑥 𝐾−1 𝑦 𝐾−1 is then interpreted as a complex number 𝑥 𝑘 +
𝑗𝑦 𝑘, for k = 0, 1, . . . , K − 1.
 “The discrete Fourier transform (DFT) of this list of complex numbers is the
Fourier descriptor of the boundary”.
 The inverse DFT restores the original boundary.
 Figure 18.14 shows a K-point digital boundary in the x-y plane and the first
two coordinate pairs, 𝑥0, 𝑦0 & 𝑥1, 𝑦1
6/3/2014 ]2ushin $hah 42

Fourier Descriptors
6/3/2014 ]2ushin $hah 43
 One of the chief advantages of
using Fourier descriptors is their
ability to represent the essence of
the corresponding boundary using
very few coefficients.
 This property is directly related to
the ability of the low-order coefficients
of the DFT to preserve the main aspects
of the boundary, while the high-order
coefficients encode the fine details.

Fourier Descriptors
 Following is a way of using the Fourier transform to analyse the shape of a
boundary.
1. The x-y coordinates of the boundary are treated as the real and
imaginary parts of a complex number
2. Then the list of coordinates is Fourier transformed using the DFT
3. The Fourier coefficients are called the Fourier descriptors.
4. The basic shape of the region is determined by the first several
coefficients, which represent lower frequencies
5. Higher frequency terms provide information on the fine detail of the
boundary
6/3/2014 ]2ushin $hah 44

Fourier Descriptors
6/3/2014 ]2ushin $hah 45

Properties of Fourier Descriptors
 Translation
 Adding some constant to values of all coordinates
 So, we only change the zero-frequency component. (Mean position only
nothing about the shape)
 So, except for the zero-frequency component, Fourier Descriptors are
translation invariant.
 Rotation
 Rotation in the complex plane by angle θ is multiplication by exp(jθ)
 So, rotation about the origin of the coordinate system only multiplies the
Fourier descriptors by exp(jθ)
6/3/2014 ]2ushin $hah 46

Properties of Fourier Descriptors
 Scaling
 It means multiplying x(k) and y(k) by some constant.
 Hence, Fourier descriptors are scaled by the same constant (Again, we
ignore the value of the zero-frequency component)
 Starting Point
 Changing starting point is equivalent to translation of the one-dimensional
signal s(k) along the k dimension
 Hence, translation in the spatial domain (in this case, k) is a phase-shift in
the transform.
 So, the magnitude part of a(u)is invariant to the start point, and the phase
part shifts accordingly
6/3/2014 ]2ushin $hah 47

HISTOGRAM-BASED (STATISTICAL) FEATURES
Histogram-basedfeaturesarealsoreferredtoasamplitudefeatures
 Histograms provide a concise and useful representation of the intensity
levels in a gray-scale image.
 The simplest histogram-based descriptor is the mean gray value of an
image, representing its average intensity m and given by
𝑚 =
𝑗=0
𝐿−1
𝑟𝑗 𝑝(𝑟𝑗)
 where 𝑟𝑗 is the jth gray level (out of a total of L possible values), whose
probability of occurrence is p(𝑟𝑗).
6/3/2014 ]2ushin $hah 49

1. The mean gray value can also be computed directly from the pixel values
from the original image f (x, y) of size M × N as follows:
𝑚 =
1
𝑀𝑁
𝑥=0
𝑀−1
𝑦=0
𝑁−1
𝑓(𝑥, 𝑦)
 The mean is a very compact descriptor (one floating-point value per image or
object) that provides a measure of the overall brightness of the corresponding
image or object.
 It is also RST invariant.
 On the negative side, it has very limited Expressiveness and Discriminative
power.
6/3/2014 ]2ushin $hah 50

2. The standard deviation(as descriptor) σ of an image is given by
𝜎 =
𝑗=0
𝐿−1
𝑟𝑗 − 𝑚
2
𝑝 𝑟𝑗
 where m is mean which define in previous slides.
 The square of the standard deviation is the variance, which is also known as
the normalized second-order moment of the image.
 The standard deviation provides a concise representation of the overall
contrast.
 Similar to the mean, it is compact and RST invariant, but has limited
expressiveness and discriminative power.
6/3/2014 ]2ushin $hah 51

3. The skew of a histogram is a measure of its asymmetry about the mean
level. It is defined as
𝑠𝑘𝑒𝑤 =
1
𝜎3
𝑗=0
𝐿−1
𝑟𝑗 − 𝑚
3
𝑝 𝑟𝑗
 where σ is the standard deviation.
 The sign of the skew indicates whether the histogram’s tail spreads to the right
(positive skew) or to the left (negative skew).
 The skew is also known as the normalized third-order moment of the image.
6/3/2014 ]2ushin $hah 52

 If the image’s mean value (m), standard deviation (σ), and mode ( defined as
the histogram’s highest peak) are known, the skew can be calculated as
follows:
𝑠𝑘𝑒𝑤 =
𝑚 − 𝑚𝑜𝑑𝑒
𝜎
4. The energy descriptor provides another measure of how the pixel values are
distributed along the gray-level range: images with a single constant value
have maximum energy (i.e., energy = 1); images with few gray levels will
have higher energy than the ones with many gray levels. The energy
descriptor can be calculated as
𝑒𝑛𝑒𝑟𝑔𝑦 =
𝑗=0
𝐿−1
𝑝 𝑟𝑗
2
6/3/2014 ]2ushin $hah 53

5. Histograms also provide information about the complexity of the image, in
the form of entropy descriptor.
“ The higher the entropy, the more complex the image”.
 Entropy and energy tend to vary inversely with one another. The
mathematical formulation for entropy is
𝑒𝑛𝑡𝑟𝑜𝑝𝑦 = −
𝑗=0
𝐿−1
𝑝 𝑟𝑗 log2 𝑝 𝑟𝑗
 Histogram-based features and their variants are usually employed as texture
descriptors, as we shall see in next slide.
6/3/2014 ]2ushin $hah 54

Texture Features
 Texture can be a powerful descriptor of an image (or one of its regions).
 Image processing techniques usually associate the notion of texture with
image (or region) properties such as Smoothness (or its opposite, roughness),
Coarseness, and Regularity.
 Figure 18.16 shows one example of each and Figure 18.17 shows their
histograms.
 There are three main approaches to describe texture properties in image
processing: Structural, Spectral, and Statistical.
 Most application focus on the statistical approaches, due to their popularity,
usefulness and ease of computing.
6/3/2014 ]2ushin $hah 55

Texture Features
 FIGURE 18.16 Example of images with smooth (a), coarse (b), and regular (c)
texture. Images from the Brodatz textures data set.
FIGURE 18.17 Histograms of images in Figure 18.16.
6/3/2014 ]2ushin $hah 56

Texture Features
• Highest uniformity has lowest entropy
6/3/2014 ]2ushin $hah 57

 Histogram-based texture descriptors are limited by the fact that the
histogram does not carry any information about the spatial relationships
among pixels. One way to circumvent this limitation consists in using an
alternative representation for the pixel values that encodes their relative
position with respect to one another.
 One such representation is the gray-level co-occurrence matrix G, defined as a
matrix whose element g(i, j) represents the number of times that pixel pairs with
intensities zi and zj occur in image f (x, y) in the position specified by an operator d.
The vector d is known as displacement vector:

Graylevelco-occurrenceMatrix
 The gray-level co-occurrence matrix can be normalized as follows:
 where Ng(i, j) is the normalized gray-level co-occurrence matrix. Since all values of
Ng(i, j) lie between 0 and 1, they can be thought of as the probability that a pair of
points satisfying d will have values (zi, zj).
 Co-occurrence matrices can be used to represent the texture properties of an
image.
 Instead of using the entire matrix, more compact descriptors are preferred.
These are the most popular texture-based features that can be computed from
a normalized gray-level co-occurrence matrix Ng(i, j):

Example
• For the given binary image compute the descriptor values
and fill the it in the given table.

Table for feature extraction results
Question 2 Do the results obtained for the extracted features correspond to your
expectations? Explain.
Question 3 Which of the extracted features have the best discriminative power to
help tell squares from circles? Explain.
Question 4 Which of the extracted features have the worst discriminative power
to help tell squares from circles? Explain.
Question 5 Which of the extracted features are ST invariant, that is, robust to
changes in size and translation? Explain.
Question 6 If you had to use only one feature to distinguish squares from circles,
in a ST-invariant way, which feature would you use? Why?

Texture Features
 One of the simplest set of statistical features for texture description consists
of the following histogram-based descriptors of the image (or region): mean,
variance (or its square root, the standard deviation), skew, energy (used as a
measure of uniformity), and entropy, all of which were introduced in Section
18.5.
 The variance is sometimes used as a normalized descriptor of roughness (R),
defined as
𝑅 = 1 −
1
1 + 𝜎2
 Where, 𝜎2
is the normalized (to a [0, 1] interval) variance.
 R = 0 for areas of constant intensity, that is, smooth texture.
6/3/2014 ]2ushin $hah 64

Image feature extraction

More Related Content

What's hot (20)

Similar to Image feature extraction (20)

More from Rushin Shah (10)

Recently uploaded (20)

Image feature extraction