Name of Student: Class:

Semester/Year: Roll No:

Date of Performance: Date of Submission:
Examined By: Experiment No: Mini Project

ASSIGNMENT – Mini Project

AIM: Mini Project Implementation

TITLE: “Feature Extraction using Zernike Moments”

OBJECTIVES:

1. To understand the concept of feature extraction and its importance in machine

learning.

2. To understand the concept of Zernike moments and their mathematical properties.

3. To implement feature extraction using Zernike moments in a programming language,

such as Python.

4. To apply the feature extraction technique to a dataset of images and analyze the
results.

5. To compare the performance of Zernike moment feature extraction with other feature
extraction techniques.

APPRATUS:
• Python Programming Knowledge.
• NLP Algorithms.
• Computer with appropriate software for data analysis and programming (e.g.,
Python)
• Dataset for classification tasks (e.g., MNIST dataset)
• A software library for image processing and analysis, such as OpenCV
• Jupyter Notebook.

ABSTRACT:

1
Zernike moments possess valuable mathematical properties that make them highly suitable as
image features for shape classification tasks. Their rotational invariance ensures that the
descriptors remain consistent regardless of the orientation of the shape. Furthermore, with
additional techniques, Zernike moments can be made scale and translational invariant,
enhancing their robustness for different image scales and positions. However, the correct
application of Zernike moments requires careful consideration of various factors. This paper
explores several techniques to optimize the usage of Zernike moments as shape descriptors,
and the experimental results have demonstrated exceptional accuracy, reaching up to 100% in
certain cases. These findings highlight the effectiveness and potential of Zernike moments in
shape classification problems.

TABLE OF CONTENTS:

Sr. No. Content Page No.

1 Abstract 2
2 Introduction 2
3 Implementation 3
4 Result and Analysis 7
5 Conclusion 8
6 References 9

INTRODUCTON:
Moments have been used in image processing and classification type problems since
Hu introduced them in his groundbreaking publication on moment invariants [4]. Hu used
geometric moments and showed that they can be made to be translation and scale invariant.
Since then more powerful moment techniques have been developed. A notable example is
Teague’s work on Zernike Moments (ZM); he was the first to use the Zernike polynomials
(ZP) as basis functions for the moments [6]. ZM’s have been used in a multitude of
applications with great success and some with 99% classification accuracy [1]. The use of
ZP’s as a basis function is theoretically beneficial because they are orthogonal polynomials
which allows for maximum separation of data points, given that it reduces information
redundancy between the moments. Their orthogonal properties make them simpler to use
during the reconstruction process as well. Furthermore, the magnitude of ZM’s are
rotationally invariant, which is crucial for certain image processing applications, such as
classifying shapes that are not aligned. The breakdown of this paper is as follows. Section 2
introduces standard geometric moments and their application in achieving rotational and
translational invariance of the images. Section 3 defines Zernike polynomials and Zernike
moments. Section 4 explains the rotational invariant properties of the magnitudes of Zernike
moments. Section 5 discusses how to achieve reconstruction from Zernike moments.
Furthermore, it describes how to use these reconstructions to apply weights on these moments

2
so as to make the moments which capture the most amount of information about the image
contribute the most amount of ’votes’ for the classification. Section 6 gives a description of
the experiments used to test the un-supervised classification accuracy of Zernike moment
feature vectors. In Section 7 we will present our conclusion of our results, and our
justification for why these results should be used in applicable image classification problems.
Finally, In Section 8 a brief overview of future research is presented.

IMPLEMENTATION:

Let’s go ahead and introduce our two images. The first image, our reference image, can be
seen below:

Here we can see a Game Boy ROM cartridge of game, Pokemon.

3
Given this reference image of the Pokemon game cartridge, our goal is to detect it in the
following image filled with distractor objects:

In this image we can see that we have the Pokemon game cartridge. But also also have a
bunch of other distractor objects, such as scissors, a highlighter, and a sticky note.

The first thing we’ll do here is define a describe_shapes function. This function will be
applied to the Pokemon game cartridge in the reference image along with every object in the
second image. Given that this section will be called so many times, I thought it best to create
a dedicated function to describing the shape regions.

 The describe_shapes function will take a single argument, an image that contains
the objects and shapes we want to quantify using Zernike Moments.

We’ll need to apply a little bit of pre-processing to our image before we can extract our
shape features, so we’ll convert the image to grayscale, blur it to remove high frequency
noise and allow us to focus on the structural aspects of the image, followed by thresholding to
segment the objects from the background.

4
We’ll also perform a series of dilations and erosions to close gaps between any parts of
objects.

5
Now that the Morphological Operations have been applied, we can then detect the contours
of each of the objects in the image.
We start looping over each of the individual contours. And then for each of these contours,
we allocate memory for a mask followed by drawing the contoured region on the mask.
Now that we have just the contoured region drawn on the mask, we can then compute the
bounding box and extract the ROI (from the mask image). Again, this step mentioned above
to ensure that only the current shape was drawn on the mask and nothing else.
Given the ROI of the shape we can now extract Zernike Moments by making a call to
the mahotas.features.zernike_moments function. This function accepts 3 parameters: the
image/ROI that we want to quantify, the radius of the region, and the degree of the
polynomial.
** Tuning the radius can be non-trivial based on the sizes of the objects in the image. If the
sizes of the objects vary dramatically, as they do in our distractor image above, the value of
this radius is often not immediately obvious.

To solve this problem, we have 2 solutions :

 The first solution is to simply resize the ROI to a known size. This ensures that we
will be able to hardcode the radius value and that our images are described in a
consistent manner.
 The problem with this approach is that resizing to a fixed size can destroy the aspect
ratio of the shape. This is especially troublesome if we are trying to quantify the
shape of a region — by throwing away the aspect ratio, we throw away information
regarding the dimensions of the shape, which is completely counterintuitive to our
goal!

The other solution is what this example utilizes — we can dynamically compute the radius of
the Zernike Moments simply by computing the cv2.minEnclosingCircle function (a simple
contour property). This method will return the minimum size radius that can be used to
enclose the entire object. By applying this method, we can ensure the radius of the moments
covers the entire ROI.

 We also supply degree=8 to the zernike_moments function, which is the default

degree of the polynomial. In most cases you’ll need to tune this value until it obtains
adequate results.

Finally, we can return a 2-tuple from our describe_shapes function consisting of (1) the
contours of the objects/shapes in the image, followed by (2) the Zernike Moments feature
vectors corresponding to each shape.
So now that we have the describe_shapes function defined, let’s see how we can use it to
recognize the Pokemon game cartridge in an image:

We load our reference image of the Pokemon game cartridge from disk and describe shape of
the game cartridge.

We load the distractor image from disk and quantify all of the shapes in the image using
Zernike Moments.

6
To detect the actual game cartridge, we use SciPy’s distance sub-module to compute the
Euclidean distance between all pairs of the gameFeatures and shapeFeatures.

 Notice how we have only one row, which is the Zernike Moments associated with the
game cartridge features. We also have multiple columns — one for each of the shapes
in the shapeFeatures list.

The index of the column that minimizes this distance is thus our Pokemon game cartridge.
How can we be sure??
Zernike Moments are used to quantify the shape of an object. If we assume shapes that
have similar feature vectors also have similar visual contents, then the shape that minimizes
the distance between Zernike Moments must be our reference image!
We’ll start by looping over the contours in our distractor image. If the index of the current
contour does not match the index of the contour with minimum distance in our distance
matrix D, then we draw a rotated bounding box surrounding it in red.
We then then draw a green rotated bounding box surrounding our identified game cartridge
region, followed by displaying the text “FOUND!” directly above the bounding box.

Result and Analysis:

On the left we have our original reference image of the Pokemon game cartridge. And on
the right we have our distractor image containing not only the Pokemon game cartridge, but
other objects and shapes meant to “confuse” our algorithm.

However, by applying Zernike Moments features we were able to discard the distractor
shapes and detect the actual game cartridge shape in the image.

7
CONCLUSION:
In conclusion, when describing multiple shapes in an image using Zernike Moments, it is
important to follow a few key steps. First, extract the region of interest (ROI) for each object
in the image. Then, apply Zernike Moments to each ROI to capture the shape characteristics.
When utilizing Zernike Moments, it is crucial to consider the radius and degree parameters.
The radius determines the size of the disc onto which the shape is mapped. It is essential to
choose a radius that is sufficiently large to include all relevant pixels of the shape. If any
pixels fall outside the radius, they will be disregarded. Therefore, careful attention must be
paid to ensure an appropriate radius is selected before extracting Zernike Moments.
Additionally, the degree parameter directly affects the dimensionality of the resulting feature
vector. Higher degree values lead to larger and potentially more distinctive feature vectors.
However, it is important to consider the computational cost as the number of degrees
increases.
When setting the radius and degree parameters, it is recommended to prioritize the radius
selection. Reflecting on the size of the objects in the dataset will help determine an adequate
radius that captures their shapes effectively. Once the radius is established, the degree
parameter can be fine-tuned accordingly. Starting with a value of degree=8 is often a
reasonable initial choice, with adjustments made as needed based on the desired
dimensionality of the feature vector.
By following these guidelines, one can effectively describe shapes in an image using Zernike
Moments, taking into account the appropriate radius and degree parameters to capture the
desired shape characteristics.

REFERENCES:
[1] Khotanzad A. and Hong Y. H. Invariant image recognition by zernike moments. IEEE,
12(5):489 – 497, 1990.
[2] Suk T. Flusser J. and Zitova B. Moments and Moment Invariants in Pattern Recognition.
Wiley and Sons Ltd., 2009.
[3] Fan X.X. Fu B., Liu J. and Quan Y. A hybrid algorithm of fast and accurate computing
zernike moments. IEEE, 2007.
[4] Hu M. K. Visual pattern recognition by moment invariants. IRE Transactions on
Information Theory, 8(2):179 – 187, 1962.

8
[5] Yu M. Feature-Weighted Hierarchical Sparse Learning for Semantic Concept
Identification.
[6] Teague M. R. Image analysis via the general theory of moments. Optical Society of
America, 70(8):920 – 930, 1979.
[7] Prokip R.J. and Reeves A.P. A survey of moment-based techniques for unoccluded object
representation and recognition. CVGIP: Graphical Models and Image Processing, 54(5):438 –
460, 1991.
[8] Witlin R.S. The witlin center for advanced eyecare, 2008.