IRJET- Semantic Retrieval of Trademarks based on Text and Images Conceptual Similarity using Deep Learning

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 12 | Dec 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 662
“Semantic Retrieval of Trademarks based on Text and Images
Conceptual Similarity using Deep Learning”
Prof. Pramod Dhamdhere1, Ashwini Nilakh2, Sushmita Choudhari3, Komal Jadhav4,
Namrata Kate5, Ankita Temgire 6
1Professor, Dept. of information Technology, BSIOTR College, Wagholi, Pune, Maharashtra, India
2,3,4,5,6UG. Students, Dept of information Technology, BSIOTR College, Wagholi, Pune, Maharashtra, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - The number of images associated with weakly
supervised user-provided tags has increased dramatically in
recent years. User-provided tags are inadequate, subjective
and noisy. In proposed system, focusedontheproblemofsocial
image understanding, i.e., tag refinement, tagassignment, and
image retrieval. Different from past work, system propose a
novel weakly supervised deep matrix factorization algorithm,
which uncovers the latent image representations and tag
representations embedded in the latent subspace by
collaboratively exploring the weakly supervised tagging
information, the visual structure, and the semantic structure.
The semantic and visual structures are jointly incorporated to
learn a semantic subspace without over-fitting the noisy,
incomplete, or subjective tags. Besides, to remove the noisy or
redundant visual features, a sparse model is imposed on the
transformation matrix of the first layer in the deep
architecture. Extensiveexperimentsonreal worldsocial image
databases are conducted on the tasksofimage understanding:
image tag refinement, assignment, andretrieval. Encouraging
results are achieved, which demonstrates the effectiveness of
the proposed method. Finally, a unified optimization problem
with a well-defined objective function is developed to
formulate the proposed problem and solved by a gradient
descent procedure with curvilinear search. Extensive
experiments on real world social image databases are
conducted on the tasks of image understanding: image tag
refinement, assignment, andretrieval. Encouragingresultsare
achieved with comparison withthestateof-the-artalgorithms,
which demonstrates the effectiveness of the proposedmethod.
A trademark is a mark that you can use to recognize your
business products or services from those of other vendors. It
can be represented graphically in the form of any Symbol,
logo, words etc. so, they need to be protection. The conceptual
similarities among trademarks, which happens when more
than two or more trademark similar.
Keywords: User Provided Tags, Image Tag Refinement,
Image Tag Assignment, Image Tag Retrieval, Social
Image Understanding.
I) INTRODUCTION
In the social media networks human is considered as open
and complex framework. The requirements of the user
changed likewise because the expectationofone person may
subspace by cooperatively investigating the weakly
supervised tagging data, semantic structure and visual
structure. Recent years have witnessed an increase in the
number of community-contributed images associated with
rich contextual information such as user-provided tags.
These users gave tags can portray the semantic substance of
pictures to some degree, which is valuable to numerous
tasks, for example, picture tagging (which can be treated as
an image to tag search), Content-Based Image Retrieval
(CBIR) and Tag-BasedImageRetrieval (TBIR).Subsequently,
it is vital yet difficult to cooperatively investigate the rich
data of network contributed pictures that is regularly
normally accessible. By and by, connections are constantly
required for numerous tasks, for example, picture tag
relationship for cross modular search (i.e., picture tagging
and TBIR), picture relationship for CBIR and tag connection
for tag extension in true applications, and these connections
must be exact. The quantity of pictures related with weakly
supervised user-provided tags has expandedsignificantlyas
of late. User-provided tags are insufficient, abstract and
boisterous. System centre on the issue of social picture
understanding, for example tag assignment, image retrieval
and tag refinement. System propose a weakly supervised
deep matrix factorization algorithm, in which reveals the
inactive picture portrayals and tag portrayalsinsertedin the
dormant environment in service is awareness about the
circumstance. Then it can be easily adjusted to the dynamic
service.
II) LITERATURE SURVEY
J. Tang et al. [5] present many image processing and pattern
recognition problems; visual contents of images are
currentlyde-scribedbyhigh-dimensional features,whichare
often repetitive and loud. Creators proposed a novel
unsupervised component choice plan, to be specific, non-
negative phantom investigation with obliged excess, by
together utilizing non-negative otherworldly clustering and
redundancy analysis. The presented method can directly
identify a discriminative subset of the most useful and
redundancy-constrained features.
Z. Li et al. [6] present performance of TBIR is limited due to
incorrect or noisy tag associated withtheimageuploaded on
social websites. To overcome the performance issues some
previous image retagging techniques are proposed to fine
tune the tag information of social image in transductive
learning manner. However, most of the techniques are
unable to handle the images which are not part of sampling
data. In author proposed an approach of novel factorization
called as Projective matrix factorization with unified

embedding for tag learning and retagging. The learning
phase previously tagging information of social images is
applied to tag correlation matrix and find image. This can
handle the large-scale social image retagging tasks.
Z. Lin et al. [7] present despite the fact that generally used
for encouraging picture the executives, client gave picture
labels are normally inadequate and deficient to portray
entire semantic substance of relating pictures, bringing
about execution debasement in label subordinate
applications and in this way requiring powerful label
consummation strategies. System proposed a novel plan
indicated as LSR for programmed picture label finishing by
means of picture explicit and tag-explicit Linear Sparse
Reconstructions. Given an inadequateintroductorylabelling
grid with each line speaking to a picture and every segment
speaking to a tag, LSR ideally reproduces each picture (for
example push) and each tag (for example section) with
staying ones under imperatives of sparsity, considering
picture likeness, picture label affiliation and tag-label
simultaneousness.
In this paper [8], The recent trademark reflow system of
working with reformed reflow execution for the unification
of global and local expositors. The global expositors are
using the Zernike moment’s coefficients and the local
expositors are the edge-gradient co-occurrence matrix,
defines as outline data that means it’s mainly significance in
human cognition of estimation equality. The defined reflow
system is tested use the standard MPEG-7 shapes. The
results reformation in the case of the MPEG-7 shape
databases. The bonding during two proximate factors is
hold on by usage the co-occurrence matrix on incline data.
The research in the round of offered a novel system for
trademark reflow that increase the execution.
Author proposed [9], A recent system for counting short-
text and sentence semantic similarity. The method is
depends on the concept that the sense of a statement is
create of nope mere the sense of its particular words, but
also the anatomical path the words are concatenated. Thus
hold on and connects syntactic concatenated. Thus hold on
and connects syntactic and semantic data to count the
semantic similarity of two phrases. Semantic data is given
from lexical resources. Syntactic data is get from a strong
parsing procedure that searches the sentences in every
phrase. A syntax-based providence tocalculatethesemantic
similarity between phrases or short texts. The concept on
which the system is based on the sense of phrases is
creating of nope mere the senses of its particular words, but
as well the different words are concatenated.
III) SYSTEM ARCHITECTURE
System proposed a novel Weakly-supervised Deep Matrix
Factorization (WDMF) algorithm for social image tag
refinement, assignment and retrieval, which uncovers the
latent image representations and tag representations
embedded in the latent subspace by collaboratively
exploiting the weakly supervised tagging information, the
visual structure and the semantic structure. The proposed
approach can deal with the noisy, incomplete or subjective
tags and the noisy or redundant visual features. The
proposed approach is formulated as a joint optimization
problem with a well-defined objective function, which is
solved by a gradient descent procedure with curvilinear
search. Extensive experiments on two real-world social
image databases are conducted to demonstrate the
effectiveness of the problem.
There are several potential researchdirections.First,system
will explore the effective of the depth in the deep MF
framework and how toadaptivelyidentify betterparameters
for different vision tasks. Second, system will explorehow to
integrate the proposed deep MF model and CNN into a
unified framework.Besides, systemwill extendthe proposed
deep MF framework to make it applicable and investigateits
new applications such as image suggestion.
IV). ALGORITHM
V) CONCLUSIONS
Proposed novel Deep CollaborativeEmbedding(DCE)model
for social image understanding. It incorporates the end-to-
end learning and cooperative calculate investigation one
brought together structure for the ideal similarity of
representation learning and inactive space revelation. To
cooperatively investigate the rich logical data of social
images, it factorizes various correlationmatricesatthesame
time and flawlessly. A refined tagging matrix with non-
negative and discrete properties is specifically figured out
how to deal with the noisy tags. The proposed strategy is
connected to social image tag refinement and assignment,
content-based image recovery, tag-based image recovery
and tag expansion. Weakly-supervised Deep Matrix
Factorization (WDMF) algorithm for social image tag

refinement, assignment and retrieval, which uncovers the
latent image representations and tag representations
embedded in the latent subspace by collaboratively
exploiting the weakly supervised tagging information, the
visual structure and the semantic structure. To well handle
the out-of-sample problem, the underlying image
representations are assumed to be progressively
transformed from the visual feature space. Besides, the
proposed approach can deal with the noisy, incomplete or
subjective tags and the noisy or redundant visual features.
The proposed problem is formulated as a joint optimization
problem with a well-defined objective function, which is
solved by a gradient descent procedure with curvilinear
search.
VI) REFERENCES
1. S. Hong, J. Choi, J. Feyereisl, B. Han, and L. S. Davis,
“Joint Image Clustering And Labelling By Matrix
Factorization”, IEEE Trans. on Pattern Analysis and
Machine Intelligence, vol. 38, no. 7, pp. 1411–1424,
2016.
2. Q. You, H. Jin, Z.Wang, C. Fang, and J. Luo, “Image
Captioning With Semantic Attention”, in Proc. of
IEEE Conf. on Computer Vision and Pattern
Recognition, 2016.
3. L. Xu et al., “Multi-Task Rank Learning for Image
Quality Assessment”, IEEE Trans. Circuits Syst.
Video Technol., 2016,
doi:10.1109/TCSVT.2016.2543099
4. F. M. Anuar, Yu-Kun Lai, R. Setchi,”Semantic
Retrieval of Trademarks Basedon Conceptual
Similarity.” IEEE TransactionsonSystems,Man,and
Cybernetics: Systems IEEE permission, 2015.
5. Z. Li and J. Tang, “Unsupervised Feature Selection
Via Non Negative Spectral AnalysisandRedundancy
Control”, IEEE Trans. Image Process., vol. 24,no.12,
pp. 53435355, Dec. 2015.
6. Z. Li, J. Liu, J. Tang, and H. Lu, “Projective Matrix
Factorization With Unified Embedding For Social
Image Tagging”, Comput. Vis. Image Understand.,
vol. 124, pp. 7178, Jul. 2014.
7. Z. Lin, G. Ding, M. Hu, J.Wang, and X. Ye, “Image Tag
Completion Via Image Specific And Tag-Specific
Linear Sparse Reconstructions”, in Proc. IEEE
Comput. Vis. Pattern Recognit., Jun. 2013, pp.
16181625.
8. H. Qi, K. Q. Li, Y. M. Shen, and W. Y.Qu, ”An effective
solution for trademark image retrieval by
combining shape description andfeaturematching,”
Pattern Recognit., vol. 43, no. 6, pp. 2017-2027,
2010.
9. J. J. Jiang and D.W. Conrath, ”Semantic similarity
based on corpus statistics and lexical taxonomy,” in
Proc. Int. Conf. Res. Comput.Linguist, Taipei,
Taiwan, pp.19-33.
10. J. P. Eakins, J. M. Boardman, and K. Shields,
“Retrieval of trademark images by shape feature—
The ARTISAN project,” in Proc. IEEEColloq. Intell.
Image Databases, London, U.K., 1996, pp. 9/1–9/6.

IRJET- Semantic Retrieval of Trademarks based on Text and Images Conceptual Similarity using Deep Learning

Recommended

More Related Content

What's hot (12)

Similar to IRJET- Semantic Retrieval of Trademarks based on Text and Images Conceptual Similarity using Deep Learning (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Semantic Retrieval of Trademarks based on Text and Images Conceptual Similarity using Deep Learning