Moment based features, Moment Invariants,

Normalized central moments and 7 moments

of Hu

Image Processing CS 8009

Presented to: Dr. Akshay Pandey

Presented by: Sumedh Atreya

22BCS267
 Agenda

1. Introduction to feature extraction

2. Moment Functions, Moments and Order of Moments

3. Types of moments in image processing analogous to statistics and mechanics.

4. Why Normalize? & 7 moments of Hu

5. Moment Based Features in digital image processing

6. Case Studies, Advantages and Drawbacks

7. Appendix
 Introduction to Feature extraction

Definition: Feature extraction is the process of identifying and quantifying meaningful

information from raw pixel data to simplify analysis.

It transforms high-dimensional images into compact, representative descriptors.

Applications: Reduce the complexity by discarding redundancy in pixel information by identifying

patterns from the data obtained.

Capture patterns like shapes, textures, or structures for interpretation.

Enable consistent recognition under transformations (e.g., rotation,translation, scaling).

Used for: Object detection,Image Classification & Segmentation,Sentiment Analysis,

Anomaly detection,Medical image analysis,etc..
Some of the key features in feature extraction

Feature Type Purpose Example

Edges,corners Boundary detection Canny edges, Harris corners

Textures Pattern recognition Gabor filters,LBP

Shape Descriptors Geometrical Contours,Moments

Feature Extraction
Characterization of Shape
One of the basic problems in the design of any pattern recognition system is the selection of a set
of appropriate features to be extracted from the object of interest/region of interest(ROI).

Describing digit images with moments instead of other more commonly used pattern recognition
features means that global properties of the digit image are used rather than local properties.

Moment Functions, Moments and order of moments

Moments functions in image processing are used for invariant pattern recognition, object
classification,etc.. A set of moments represent global characteristics providing information on
geometrical features or the shape descriptors of an image.

In image processing, moment functions can provide information such as total image area,
coordinates of the image centroid and the orientation which can further be used to describe the
feature vectors that are invariant w.r.t transformations like translation,scaling and rotation.
Moment Functions, Moments and order of moments
Moments of different orders are used to characterize and analyze probability distributions and
stochastic processes, offering insights beyond the basic mean and variance. Higher-order
moments provide a more detailed picture of the distribution's shape, symmetry, and behavior.

Moments are helpful for determining summary measures of probability distributions. In the context
of our subject, these distributions refer to the density of pixel intensity values. We will discuss two
types of moments.
 Types of Moments

1. Geometrical moments or Moments about the origin

Geometric moments are defined as the projection of the image intensity function f(x, y) onto the
monomial xpyq ,the basis set being {xp,yq} .The (p + q)th order geometric moment Mpq of a gray
level image f(x, y) is defined as

For that of a digital image to select a meaningful subset of the moment values that contain
sufficient information to characterize the image uniquely for a specific application the equation is
replaced by:

Where I(x,y) or f(x,y) denote the intensity at pixel location (x,y) where x,y are monomials of (p+q)th
order of moment and image can be thought of as a 2D density distribution f(x,y.
Types of Moments
While raw moments are not inherently invariant, they can be used to calculate central moments
which are invariant to translation of the object since they are defined about the COM of the object.

Central moments are in turn used to describe and define moment invariants which are used as
features describing the shape or the geometrical aspects of an object.

Simple image properties derived via raw moments include:

Area(for binary images) or sum of grey level(for greytone images): M00

Centroid: {x, y}={ M10/M00 , M01/M00 }
2. Central Moments
When the f(x,y) changes under transformations like translation, rotation and scaling then the image
may be positioned such that the COM coincides with the origin of the respective field of view or
simply put,they are moments calculated w.r.t objectʼs centroid, denoted by µpq.
Types of Moments/Central moments

Where and which are obtained from raw moments.

If f(x,y) is a digital image the equation becomes,

Since central moments are computed using COM, they are equivalent to mpq whose center is
shifted to centroid of the image, henceforth it is translation invariant.

Why Normalise?
We can obtain scale invariance by normalizing these central moments.
Normalized central moments are denoted by ηpq .

In statistics, normalized central moments are typically central moments divided by a power of the
standard deviation(σ) making them dimensionless, enabling us to compare different distribution
functions regardless of their scale.
Types of Moments/Normalized Central moments and Hu invariants

The (p+q)th normalized central moment is defined as,

From these normalized central moments, we can calculate the seven derived quantities attributed
to Hu which are invariant to most of the transformations like scaling, translation and rotation.
Moment Based Features

Mathematically, moments are used to describe quantities at a distance from a reference axis/point.

Moments in statistics are used to characterize the distribution of random variables and similarly, in
mechanics to characterize bodies by their mass in a spatial distribution.

In image processing if we consider a binary or grey level image projected as a 2D DDF by their
intensity values, moments can be used to characterize an image and extract properties that have
analogies in statistics or mechanics. These properties can be said to as moment based features.

Suppose we have a region R containing N pixels,

1. The Center of Mass(COM) as discussed earlier would be and

2. Orientation(𝜽): Defined as the angle of axis with the least moment of inertia. It also describes
the orientation of the principal axes.
Moment Based Features

3. Best Fit Ellipse is the ellipse whose second order of moment ,i.e. the moment of Inertia is the same
as that of the object.
Given by,
𝑰min=(𝝅/𝟒)ab3

𝑰max=(𝝅/𝟒)a3b

where a,b are the length of semi-major and semi-minor axis

respectively.

4. Principal Axes obtained from the second order moments are

described as pair of axes about which there are minimum and
maximum second order moment (moment of inertia) along the Image ellipse of a shuttle silhouette
minor and major axes respectively.

5. Radii of Gyration is another property that can be obtained from second order moment, described
about an axis is the distance from axis to a line where all of the mass may be concentrated without
changing second order along that axis. Mathematically,
Moment Based Features/Radii of Gyration

Where ROG and ROG about the x and y axes respectively defined in terms of moments.
x y

The ROG about the origin ROGCOM)is the radius of circle (h,k)0,0) where all mass may be
concentrated without change in the second order of moment about origin.

This ROG about origin ROGCOM) is invariant to image orientation and is used as a rotational invariant
feature in object representation methods.
Moment Based Features

6. Features obtained from higher order moments are Projection Skewness and Projection Kurtosis

Skewness is the measure of the degree of deviation from symmetry(about mean) of a distribution.
Kurtosis is a measurement of peakedness of a distribution.

Given by,

and

Skx and Sky indicate which side of axis is the projection skewed towards but it fails to identify if an
object is symmetric or not.

kx and ky if zero is called a gaussian distribution. Values 0 indicate flatter, less peaked distribution
while that 0 indicate narrow, more peaked distribution.

Other features might include quantities such as eccentricity, total sum of pixel intensities, orthogonal
moments such as the zernike moments, legendre moments, affine moments, complex moments, etc…
Applications

Moment based features are useful in applications like pattern characterization and recognition,object
identification, 3D object pose estimation, robot sensing, image coding and reconstruction, etc..
independent of position, size and orientation.

Character recognition Finger-vein based identification

Medical Image analysis Object identification/Target Detection

Pros and Cons of Hu moments

Pros:

● Very fast to compute.

● Low dimensional.
● Good at describing simple shapes.
● No parameters to tune.
● Invariant to changes in rotation, reflection, and scale.
● Translation invariance is obtained by using a tight cropping of the object to be described.

Cons:

● Requires a very precise segmentation of the object to be described, which is often hard in the
real world.
● Normally only used for simple 2D shapes, as shapes become complex, Hu Moments are not
often used.
● Hu Moment calculations are based on the initial centroid computation — if the initial centroid
cannot be repeated for similar shapes, then Hu Moments will not obtain good matching
accuracy.
Appendix

References
https://cvexplained.wordpress.com/2020/07/21/10-4-hu-moments/

https://iopscience.iop.org/article/10.1088/1742-6596/1962/1/012028/pdf

https://www.semanticscholar.org/paper/Robust-matching-of-fingerprints-using-moments-Kahyaei-Moin/76831b2595b8896
dc8b27822fa1383506c14d566

https://support.ptc.com/help/mathcad/r10.0/en/index.html#page/PTC_Mathcad_Help/moment_invariants.html

https://onlinelibrary.wiley.com/doi/full/10.1155/2016/2796863

https://en.wikipedia.org/wiki/Image_moment

https://davuniversity.org/images/files/study-material/Moment-Based%20Feature.pdf

https://stackoverflow.com/questions/11379947/meaning-of-the-seven-hu-invariant-moments-function-from-opencv

17
Thank You

18