Efficient Image Retrieval by Multi-view Alignment Technique with Non Negative Matrix Factorization

International Journal of Research and Scientific Innovation (IJRSI) | Volume IV, Issue IV, April 2017 | ISSN 2321–2705
www.rsisinternational.org Page 52
Efficient Image Retrieval by Multi-view Alignment
Technique with Non Negative Matrix Factorization
Chandranayaka I R, Radhakrishna M, Dipak cumar V C
Assistant Professor, Department of ECE, Global Academy of Technology, Bengaluru-98
Abstract-The biggest challenge in today’s world is a searching of
images in a large database. For searching of an image a
technique which can be termed as Hashing is used. Already there
are many hashing techniques are present for retrieval of an
image in a large databank. The hashing technique can be done
on images by considering the high dimensional descriptor of an
image but in the existing hashing techniques single dimensional
descriptor is used from this the performance of the probability of
distribution of an image search will not achieve as expected. And
the drawback is that giving the query input in a texture format
leads to the limitations of image search that is firstly due to the
limited keyword and the second is Annotation approach by
human is ambiguous and incomplete.
To overcome from these drawbacks a new technique has been
proposed named as Multiview Alignment Hashing technique in
which it keeps the high dimensional feature descriptor data as
well probability of distribution of an images in a database. Along
with the Multiview feature descriptor another technique can be
used that is Nonnegetive Matrix Factorization (NMF). NMF is a
popular technique used in data mining in which clustering of a
data will be takes place by considering only the nonnegative
matrix value.
Keywords: Image retrieval, NMF, Feature descriptor
I. INTRODUCTION
ash function to embed high-dimensional data into a
similarity-preserving low-dimensional Hamming space
such that an approximate nearest neighbor of a given query
can be found with sub-linear time complexity. Traditional
method on hashing is to compute a hash function purely in
randomized manner, where random projections can be
achieved by rounding up of the value which is used to
generate binary codes. Since it does not match with the
expected performance because of shorting of a binary codes
are used multiple hash tables. The hash technique in image
processing is the study of similarity-preserving binary codes,
which can be classified as unsupervised and semi- supervised
prototypes. Example for unsupervised hashing method is
Spectral hashing.
In the classic approach of the hashing technique is subjected
to single-view. In their system architectures it focuses mainly
on only one type of feature descriptor is used to study hashing
functions. In practical, to make a more complete feature
description, objects or images are always represented by
several different types of features and each one of the hashing
function can be defined with its own characteristics. Thus, it is
necessary to assimilate heterogeneous feature descriptors into
hashing functions, leading to multi-view hashing approaches.
Unlike other methods like with positive and negative
values, NMF considered only nonnegative parts-based
representation which provides a better visual elucidation of
factoring matrices for high-dimensional data. Therefore, in
many cases, NMF may be more suitable for subspace tasks,
since it provides a basis for a non-global set which
subconsciously contains the localized parts of objects. In
addition to this flexibility of matrix factorization can be
handling by varying data distributions. It also enables more
robust subspace and more importantly, NMF decomposes an
original matrix into a part-based representation by this way it
gives better interpretation of factoring matrices for non-
negative data.
On applying NMF technique to combing of a multi view
feature, a part-based representation can diminish the collision
between any two views and it achieves more discriminative
codes. It is worthy to highlight several contributions of the
proposed method MAH with RKNMF can find a compact
representation uncovering the hidden semantics from different
view aspects and simultaneously respecting the joint
probability distribution of data. RKNMF handles non convex
objective function.
II. REVIEWS OF EXISTING SYSTEMS
In [1], the capacity of fast resemblance search in a significant
dataset is of great prominence to many hypermedia
applications. One of the hashing technique named as Semantic
hashing is a promising way to speed up likeness search, which
can be applied for designs that contains dense binary codes
for a large number of images they are semantically related
images are mapped to close codes. In spectral hashing images
related with semantically consistent graph approach for image
indexing. Rescuing like neighbour is then simply achieved by
retrieving images that have codes within a small Hamming
distance of the code of the query. Among various hashing
approaches, spectral hashing (SH) has shown promising
performance by learning the binary codes with a spectral
graph partitioning method. The spectral hashing technique
suffers from noise to data sensitive.
In [2], hashing method is used to generate hash function to
insert high-dimensional data into a similarity-preserving low-
dimensional equal space between neighbours in an image such
H

that an approximate nearest neighbour of a given query can be
found with sub-linear time complexity. In this Multi-View
Anchor Graph Hashing (MVAGH), where non-linear
integrated binary codes are efficiently determined by a sub-set
of eigen vectors of an averaged similarity matrix. This method
suffers from the performance degradation when the high recall
is required. To overcome this drawback, simple heuristic to
combine MVAGH with locality sensitive hashing (LSH) has
been proposed. For single-view hashing methods, multi-view
data are concatenated into single representation. As a
performance measure, hamming ranking is a key term
hamming ranking exits between a query and data points are
decided by hamming distance.
In [3], hashing process is used to seek an embedding of high-
dimensional objects into a similarity-preserving low
dimensional Hamming space such that similar objects are
indexed by binary codes with small Hamming distances. A
variety of hashing methods have been developed, but most of
them resort to a single view (representation) of data. In the
proposed method deep network for multi-view hashing,
referred to as deep multi-view hashing, where each layer of
hidden nodes is composed of view-specific and shared hidden
nodes, in order to learn individual and shared hidden spaces
from multiple views of data.
In [4], the process called as Nonnegative matrix factorization
is a method used to find a low rank approximation of a
multivariate data the main objective of the Nonnegative
matrix factorization is to estimate the data matrix in a given
data. There are numerous problems that involve finding
similar items. These problems are often solved by finding the
nearest neighbour to an object in some metric space. This is
an easy problem to state, but when the database is large and
the objects are complicated, the processing time grows
linearly with the number of items and the complexity of the
object.
In [5], semantic hashing seeks compact binary codes of
data-points so that the Hamming distance between code words
correlates with semantic similarity. The proposed method
shows that the problem of finding a best code for a given
dataset is closely related to the problem of graph partitioning
and can be shown to be NP hard. By relaxing the original
problem, obtain a spectral method whose solutions are simply
a subset of threshold eigen- vectors of the graph Laplacian.
III. DESIGN OF THE PROPOSED SYSTEM
This section represents the workflow of Multiview
Alignment Hashing approach, referred as MAH. The aim of
MAH is to achieve a hash function embedding, which is used
to incorporate the various configuration representations from
heterogeneous sources. Preserving the high-dimensional
feature joint distribution is obtained and it is on the orthogonal
bases. Initially, the goal is to find the binary solution that is
converting RGB model of an image into Grey format and then
representing in a binary matrix form. However, is first relaxed
to a real valued range so that a more suitable solution can be
obtained. After applying the alternate optimization method
conversion from the real-valued solutions into binary codes
will be done and it can be stored in a database [6].
Figure.1: Architecture diagram of Multi-view Alignment Hashing
In Multiview Alignment Hashing the multi feature of an
image can be extracted including the orientation and the angle
of an image is inclined. The work flow of MAH can be
illustrated by considering the below figure.
Figure. 2: System Flow Chart
A. Training Phase
In the training phase it contains set of images to be trained.
To define this have to construct the corresponding N × N
matrix using the Heat Kernel formulation by using related
scalable parameter (τi). The approach is designed such that it
can suitable for any genuine kernel without loss of image
property. One of the most popular kernels that is Heat Kernel
is used. Heat Kernel can be computed by using multiple
kernel matrices from each view data are computed. To get a
significant low-dimensional matrix factorization, set a
constraint for the binary representation as the similarity
probability regularization, which is utilized to preserve the
data distribution in the intrinsic objective space.
B. Clustering of Image
K-means clustering is a method of vector quantization,
originally from signal processing, that is popular for cluster
analysis in data mining. K-means clustering aims to partition
n observations into k clusters in which each observation

belongs to the cluster with the nearest mean, serving as a
prototype of the cluster. This results in a partitioning of the
data space into Coronoid cells.
C. Feature Extraction
Feature Extraction is a technique of transforming arbitrary
data, such as text or images into numerical features usable for
machine learning. The latter is a machine learning technique
applied on these features. In particular in a supervised setting
it can be successfully combined with fast and scalable linear
models to train document classifiers. Classification of images
using sparse features includes an unsupervised setting it can
be used to group similar images together by applying
clustering algorithms such as K-means, Clustering text
documents using kmeans.
D. MultiviewAlignmet Hashing Embedding
The proposed method defines an embedding method. In
practical, mapping of all the training and test data in a same
way to ensure that they are in the same subspace that is they
are maintaing Hamming space which are obtained via
multiview RKNMF method. This procedure is similar as a
handcrafted embedding scheme, without retraining.
IV. EVALUATION PHASE
Figure. 3: Image Retrival on comparing with a database
The above figure summerises the Testing Phase of the
MAH method in which user will give the Query image as a
input for which the similarity check has to be done. The
multifeaure extraction of a query image is takes place that is
converting an image into numeric value understandable my
machine with an inclined orientation. In the next section the
probabiliity of similarity check is performed on comparing
with a Nearest Neighbour (NN) [7] points in a trained images
which is stored in a database. If Nearest Neighbour points are
matched then the corresponding images will be retrived and
displayed.
V. COMPLEXITY ANALYSIS
The analysis of complexity mainly contains two parts. The
first part is for the constructions of Heat Kernels and
similarity probability regularization items for different views.
And, the updating of kernel weight has the complexity of
O(n2N2) in MAH. In total, the time complexity of MAH
learning is O(2(n i=1 Di)N2 + T × (N2d + n2N2)), whereT is
the number of iterations. Stastitically, T is always less than 10,
i.e., MAH converges within 10 rounds.
VI. ADVANTAGES
This is the first work using NMF to combine multiple views
for image hashing. MAH can find a compact representation
uncovering the hidden semantics from different view aspects
and simultaneously respecting the joint probability
distribution of data. It also solve nonconvex objective
function, optimization method has been proposed to get the
final solution. By using a multivariable logistic regression to
generate the hashing function is achieved. And also it requires
less computational complexity is less as well as the data to the
noise also very less compared to the existing system methods.
VII. CONCLUSION
By using unsupervised hashing method a new novel method
has been proposed is termed as Multiview Alignment Hashing
(MAH), where hashing functions are generated by using
kernelized Nonnegative Matrix Factorization including the
preservation of a data by joint probability distribution. In
MAH implantation of multiple visual features from different
views together and an alternate way is introduced to optimize
the weights for different views and simultaneously produce
the low dimensional representation. This is addressed as a
nonconvex optimization problem and it can be overtaken on
applying locally optimal solution. For the out-of-sample
extension, multivariable logistic regression has been
successfully applied to obtain the regression matrix for fast
hash encoding.
REFERENCES
[1]. L. Shao, L. Liu, and X. Li, “Feature learning for image
classification via multiobjective genetic programming,” IEEE
Trans. Neural Netw. Learn. Syst., vol. 25, no. 7, pp. 1359–1371,
Jul. 2014.
[2]. Zhu and L. Shao, “Weakly-supervised cross-domain dictionary
learning for visual recognition,” Int. J. Comput. Vis., vol. 109,
nos. 1–2, pp. 42–59, Aug. 2014.
[3]. J. Han et al., “Representing and retrieving video shots in human-
centric brain imaging space,” IEEE Trans. Image Process., vol. 22,
no. 7, pp. 2723–2736, Jul. 2013.
[4]. J. Han, K. N. Ngan, M. Li, and H.-J. Zhang, “A memory learning
framework for effective image retrieval,” IEEE Trans. Image
Process., vol. 14, no. 4, pp. 511–524, Apr. 2005.
[5]. Kunal Shriwas and Vaqar Ansari, “Content based Image Retrieval
using Model Approach”, International Journal of Applied
Information Systems, Vol. 10,No.08, pp. 27-32, 2016.
[6]. Zhu and L. Shao, “Weakly-supervised cross-domain dictionary
learning for visual recognition,” Int. J. Comput. Vis., Vol. 109,
Nos. 1–2, pp. 42–59, Aug. 2014.
[7]. Sandeep Singh and Er.Rachna Rajput, “Content Based Image
Retrieval using SVM, NN and KNN Classification”, International

Journal of Advanced Research in Computer and Communication
Engineering (IJARCCE), Vol. 4, Issue 6, pp.549-552, 2016.
[8]. Manoharan Subramanian and Sathappan Sathappan , “An Efficient
Content Based Image Retrieval usingAdvanced Filter
Approaches”, International Arab Journal of Information
Technology, Vol.12,No.3, pp.229-236, 2015.
[9]. Vuppala swathi and M. Showkatath Ali, “Feature Based Image
Retrieval System using DWT and SVD”, International Journal of
Scientific Engineering and Technology Research, Vol.4, Issue 35,
pp.7196-7198, 2015.
[10]. K. Haridas and Dr. Antony Selvadoss Thanamani, “Well-
Organized Content based Image Retrieval System in RGB Color
Histogram, Tamura Texture and Gabor Feature”, International
Journal of Advanced Research in Computer and Communication
Engineering, Vol. 4, Issue 10, pp. 8242-8248, 2014.
[11]. Duraisamy Yuvaraj and Shanmugasundaram Hariharan, “Content-
Based Image Retrieval Based on Integrating Region Segmentation
and ColourHistogram”, The International Arab Journal of
Information Technology, Vol. 13, No.1A, pp. 203-207, 2016.

Efficient Image Retrieval by Multi-view Alignment Technique with Non Negative Matrix Factorization

More Related Content

What's hot (20)

Similar to Efficient Image Retrieval by Multi-view Alignment Technique with Non Negative Matrix Factorization (20)

More from RSIS International (20)

Recently uploaded (20)

Efficient Image Retrieval by Multi-view Alignment Technique with Non Negative Matrix Factorization