International Journal of Engineering Trends and Technology (IJETT) – Volume 22 Number 10 - April 2015
Enhanced Shape Retrieval Method Using Neural
Network Based Classification
Soma Akhil,Soma Anvesh
Soma Akhil,Accenture,India
Soma Anvesh,Tcs,India
G-15,Krishna Arcade,Nizampet,Kukatpally,Hyderabad,Telangana,India-500090
Abstract— There is a lot of research going on the images to
analyse using features such as shape, colour and texture. This
paper is going to focus mainly on 2-dimensional binary images
where shape is the only feature. Shape being the most important
feature of an image, there are many ways to retrieve the shape
features namely Moments, Shape-scale methods, Shape
transform methods. The basic aim of this paper is to retrieve
similar images from a database by giving a query image as input
using shape retrieval techniques and neural networks for 2dimensional binary images. This paper combines existing
techniques for getting better results in identifying images most
similar to query image in the database.
Keywords- Moment, Zernike moments, Legendre moments, Hu
moments, Neural networks, Re-ranking, Back propagation, Training,
database.
to describe objects after dividing them into segments. Some
properties of the image which are found via image moments include
area and centroid information about its orientation. They belong to
the region based shape invariants.
A. Zernike Moments
Zernike introduced a set of complex polynomials which forms a
complete orthogonal set over the interior unit of circle, i.e.
. These descriptors are derived using the Zernike
Moment functions. First, the 2D-image is mapped onto a unit circle
and the moment functions are applied on the circle. Since, the circle
is translation, rotation and scale invariant, the descriptors value will
be almost near to the image which we are going to match with the
query image. These are the most powerful among all the descriptors.
B. Legendre Moments
I.
INTRODUCTION
In the recent days, one of the most important task for
machines is to identify some objects of an image and retrieve similar
images by querying the image itself or using the visual features
extracted from the image. As the databases are growing large and
more and more visual features are coming into light, it is becoming
harder in extracting the features and also query similar images among
the huge databases.
The Legendre polynomials are a complete orthogonal basis
set defined over the interval [-1, 1]. The Legendre polynomials,
sometimes called Legendre functions of the first kind, Legendre
coefficients, or zonal harmonics, are solutions to the Legendre
differential equation. If l is an integer, they are polynomials. The
Legendre
polynomials
are
illustrated
above
for
and n = 1, 2,.,5.
So there is a need for new algorithms and methods for
extracting the visual features and also new search implementations
for retrieving similar images from database. In this paper, we are
implementing the latest methods for extracting the visual features
from images and also for querying the databases.
C.
In this paper, we are going to use Moment invariants for
retrieving shape features from the database images and we train the
neural network with shape descriptors of images as inputs and the
category they belong to as outputs. Once the neural network is
trained, we extract shape descriptors from the query image and fed as
input to the neural network and the output obtained will be the
category it belongs to. We apply re-ranking algorithms to the image
descriptors of the images in the output category and retrieve the most
similar images.
II.
Hu Moments
Normalized central moments are invariant to scale change.
Based on the normalized central moments, Hu introduced seven nonlinear functions. Hu’s seven moment invariants are invariant to
image scaling, translation and rotation. They have an advantage that
they can be used for disjoint shapes and disadvantages of high
computation cost, the recovery of image is difficult and computation
expensive.
Neural Networks
Moment Invariants
Basically in image processing, an image moment is defined as
a particular weighted average (moment) of the image pixels'
intensities, or a function of such moments, usually chosen to have
some interpretation or attractive property. Image moments are useful
ISSN: 2231-5381
Generally in computer science and its related fields, artificial
neural networks are computational models influenced by an
animal's central nervous systems which are capable of machine
learning as well as pattern recognition. Artificial neural networks are
http://www.ijettjournal.org
Page 467
International Journal of Engineering Trends and Technology (IJETT) – Volume 22 Number 10 - April 2015
generally shown as systems of interconnected "neurons" which can
compute values from inputs.
Let us consider a neural network for recognizing handwriting
is defined by a set of input neurons may be turned on by the pixels of
an input image. The operations of these neurons are then passed on,
transformed and weighted by a function which was determined by
the designer, to other neurons. This process is repeated many times
till an output neuron is turned on. This determines which character
was read.
Windows XP or 7 operating system with mat lab 7 or higher versions
installed. This minimum operating requirements are used for working
on the method i.e., to find similar images using shape retrieval.
C. Assumptions and Dependencies
In this algorithm we develop, we expect user to provide a
folder containing 2-dimensional binary images arranged in
category wise to extract the visual features and to train the
neural network and provide a query image of the same manner
explained above. This project expects mat lab installed on the
system and nntool to be present in the mat lab to work on the
neural networks.
Like other machine learning methods, neural networks helps to
solve many tasks that are very hard to solve using general rule-based
programming, including recognition of speech and computer vision.
D. Back Propagation Algorithm
D.
Neural back propagation is the fact in which the action
possibilities of a neuron creates a voltage increase both at the end of
the axon (normal propagation) and back through to the dendritic
arbor or dendrites, from which much of the original input current
originated. It has been shown that this simple process can be used in
a manner similar to the back propagation algorithm used
in multilayer perceptrons, a type of artificial neural network. In
addition to active back propagation of the action potential, there is
also passive electronic spread.
User Requirements
The requirements that we get from the user are the database
folder which contains 2-dimensional binary images arranged as
category wise in sub folders and query image is also a 2dimnesional binary image.
IV.
Design Overview
E. Re-ranking Algorithm
In content-based image retrieval (CBIR) systems, it was observed
that images dissimilar in visually to a query image are ranked high in
retrieval results, which affects the retrieval effectiveness. To set right
this problem, re-rank the retrieved images via clustering and
relevance feedback. Based on conventional CBIR system, the
retrieved images are analyzed using clustering method, and the
weights of each feature component are updated. Then, the rank of the
results is noted according to the distance of a cluster from a query.
Experimental results show that re-ranking algorithm achieves a more
rational ranking of retrieval results compared with existing methods.
III.
Gaps Identified from the existing systems
The existing systems work well but a combination of these
techniques provide better results. The content based image retrieval
techniques some algorithms to perform retrieval operations. That
involves a lot of computation wastage and time. Instead this project
uses a neural network for categorizing the images and predicting the
category. This decreases a lot of computation cost and also it gives
better results.
A. Proposed Solution
The existing problem of retrieving similar images needs
better shape descriptors and better prediction techniques for
retrieving the images. This can be achieved by applying moment
invariants for shape extracting the shape descriptors from the images
which are more powerful than other features. The descriptors
obtained will be shape, scale and rotation invariant. Neural networks
can do better prediction than CBIR algorithms, whereas neural
networks with re-ranking algorithms fetch better results.
B. Operating Requirements
In the process of working on the method we are in need of
some basic operating requirements such as Pentium processor, 1 GB
and above Ram with minimum hard disk capacity of 40 GB and also
ISSN: 2231-5381
http://www.ijettjournal.org
Page 468
International Journal of Engineering Trends and Technology (IJETT) – Volume 22 Number 10 - April 2015
A. System Architecture Design
V.
The projects architecture is divided into 3 phases. In the first
phase, the database will be chosen and loaded. The databases
contains 2-dimensional binary images divided into categories in the
form of sub folders. Each sub folder contains a single category. Then
it starts extracting visual features namely Zernike, Legendre and Hu
descriptors from each image and store in the form of a matrix as
inputs and category as outputs. Once visual features are extracted
from all the images in the database, then it starts training the neural
network by giving the visual descriptors as inputs and category
matrix as outputs. The neural network uses a back propagation
algorithm to increase the accuracy of the results. Once the neural
network is trained, query image is taken as input. Visual features are
extracted from the query image and fed as inputs to the neural
network. The output will be the predicted category of the query
image given. In the final phase, once the category is identified, re
ranking algorithms will be used over the image descriptor values of
all the images in that category. The filtered images will be the most
similar images and these images are displayed as outputs.
B.
Algorithm
Whole data base of images are classified into groups based
on the shape of the 2-dimensional binary images
Image descriptor values will be calculated for all the
images and the values, images are used as inputs to train
the neural network.
For the given query image all the Legendre, Zernike, Hu
moment values are calculated and moment vector is
formed.
By passing the moment vector to the neural network will
predict the category to which query image belongs.
With category know re-ranking algorithm is used to find
the images which all similar to the query image based on
the Euclidian distance.
Out of all the resultant similar images top 5 images will be
displayed.
Alternative Designs:
A. Considered Design Constraints

Initial idea was to form a vector from different kinds of
image descriptors to be applied on the images. Later forming
the same vector for the input query image and comparing the
query vector with all the vectors of the data base images.
The image with the closest vector is our required image.
Closeness is found using Euclidian distance.
Only the 2-dimensional binary images are considered, shape
is the only parameter.
A data base which has been already categorised can be
considered in order to train the neural network.
With the day by day increase in the size of image data base,
there will be a overhead of neural network training each and
everytime the data base has been updated.
VI.
Implementation: Tools used :
A. Matlab:
MATLAB (matrix laboratory) is a multi-paradigm numerical
computing environment and fourth-generation programming
language. Developed by MathWorks, MATLAB allows
matrix manipulations, plotting of functions and data, implementation
of algorithms, creation of user interfaces, and interfacing with
programs written in other languages, including C, C++, Java,
and Fortran. Although MATLAB is intended primarily for numerical
computing, an optional toolbox uses the MuPAD symbolic engine,
allowing access to symbolic computing capabilities. An additional
package, Simulink, adds graphical multi-domain simulation
and Model-Based Design for dynamic and embedded systems.
B. Nntool

As the above process is taking more computational time. Reranking algorithm helps to find the images with more
closeness values of the query vector with all the data base
images.

Still the computational time being a overhead, neural
networks are used to speed the performance and over all
classification of the image data base. Now with the less
overhead the performance of the whole process is increased
and better results are obtained.
ISSN: 2231-5381
Opens the Network/Data Manager window, which allows you to
import, create, use, and export neural networks and data.
http://www.ijettjournal.org
Page 469
International Journal of Engineering Trends and Technology (IJETT) – Volume 22 Number 10 - April 2015
Selection of the data base for training the neural network.
Displaying similar images from data base images.
VII.
Results
The mentioned algorithm generates the similar images for a given
query image from a large set of images data base. Also the image
descriptors which are information preservative, scale and translation
rotation invariants, also low computational cost. Also neural network
enhances the performance of the system.
Neural network after training with given data base images
A. Performance Analysis
By only using the image descriptors alone the computational
time is very high, as to calculate the Euclidian distance for the query
image for all the data base images descriptor values. And also using
the neural networks alone the training of the whole data base images
without image descriptors also takes lot of computational time. So
with both the techniques combined the overall performance of the
similar images finding can be speeded up.
VIII.
Selection of query image
ISSN: 2231-5381
Conclusions
By following the above procedure, we are able to generate
powerul image descriptors which gives high performance in
distinguishing different images, finding similar images and more
accurac. We are also able to scan the database faster for finding the
images similar to the query image. In this way, we are able to
provide an algorithm for faster retrieval of similar images for a query
image using powerful and efficient techniques.
http://www.ijettjournal.org
Page 470
International Journal of Engineering Trends and Technology (IJETT) – Volume 22 Number 10 - April 2015
IX.
Acknowledgement:
We would like to thank our professor Prof. Vijayarajan V for
giving us this wonderful opportunity to show our skills and creativity
in the field of Image Processing and helping with all the stumbles
and problems we faced in making this. We have put in a great effort
into this paper and implemented it with help of material and qualified
friends who have all shown their good nature and made this a success.
We also give our word as in putting our complete commitment into
doing anything of sorts and keenly look forward to working on the
techniques and algorithms in the image processing field and teaming
up with the same group to emphasize and dream of making wonders
in this field.
X.
References
[1] Mikolajczyk, Krystian, and Cordelia Schmid. "A performance
evaluation of local descriptors." Pattern Analysis and Machine
Intelligence, IEEE Transactions on 27.10 (2005): 1615-1630.
[2] Lu, Guojun, and Atul Sajjanhar. "Region-based shape
representation and similarity measure suitable for content-based
image retrieval." Multimedia Systems 7.2 (1999): 165-174.
[3] Mukundan, R., and K. R. Ramakrishnan. "Fast computation of
Legendre and Zernike moments." Pattern Recognition 28.9 (1995):
1433-1442.
[4] Kim, Whoi-Yul, and Yong-Sung Kim. "A region-based shape
descriptor using Zernike moments." Signal Processing: Image
Communication 16.1 (2000): 95-102.
[5] Teh, C-H., and Roland T. Chin. "On image analysis by the
methods of moments." Pattern Analysis and Machine Intelligence,
IEEE Transactions on 10.4 (1988): 496-513.
[6] Chong, Chee-Way, Paramesran Raveendran, and Ramakrishnan
Mukundan. "Translation and scale invariants of Legendre moments."
Pattern Recognition 37.1 (2004): 119-129.
[7] Pedronette, Daniel Carlos Guimarães, and Ricardo da S. Torres.
"Exploiting contextual information for image re-ranking." Progress
in Pattern Recognition, Image Analysis, Computer Vision, and
Applications. Springer Berlin Heidelberg, 2010. 541-548.
[8] Rahmani, Rouhollah, et al. "Localized content based image
retrieval." Proceedings of the 7th ACM SIGMM international
workshop on Multimedia information retrieval. ACM, 2005.
[9] Park, Gunhan, Yunju Baek, and Heung-Kyu Lee. "Re-ranking
algorithm using post-retrieval clustering for content-based image
retrieval." Information processing & management 41.2 (2005): 177194.
[10] Goh, A. T. C. "Back-propagation neural networks for modeling
complex systems." Artificial Intelligence in Engineering 9.3 (1995):
143-151.
ISSN: 2231-5381
http://www.ijettjournal.org
Page 471