Вы находитесь на странице: 1из 7

Semantic Image Retrieval using natural

language processing
Abstract: description of the images and return all the
Images are major paternity of content on the
images whose description is the possible
internet and the content is increasing rapidly due to the
advancement in this area. Image analysis and retrieval match. The accuracy of keyword based
is one of the active research field and researchers from image retrieval systems is far from perfect
the last decagon have proposed many efficient
because of the following reasons. If the user
approaches for the same. Semantic technologies offers
promising approach to image retrieval as it tries to map made spell mistake while describing the
the low level image features to high level. In this paper, image, this image will never be listed in the
we have proposed Semantic Image Retrieval efficient
result because of this mistake. Sometimes
and effective system is proposed. This paper presents an
application of k-means clustering algorithm to image the user has to specify the image description
Retrieval system. Semantic Image Retrieval is based on / keywords in natural language which makes
hybrid approach and uses shape, color and texture
it hard to describe the image as the user has
based approaches for classification purpose. The user
can give concept / keyword as text input or can input little ability about the natural language. It is
the image itself.CBIR is used for image retrieval. NLP very difficult to find appropriate keywords
extracts the keyword from user given query text. Based
for image description (i.e. synonym plays
on extracted keyword, similarity is measured by jaccard
similarity coefficient measure. Experimental results important role in image retrieval). Content
show the efficiency / accuracy of the proposed system based Image retrieval (CBIR) has been
and support the implementation of the same.
studied for many years which locus on
Keywords: extracting and comparing features from the
ImageRetrieval, Semantic Image; Semantic Retrieval,
natural language processing, support vector machine. images [20, 21, and 22]. Image Features are
usually distill using dominant color,
1. INTRODUCTION dominant texture, or shape Researchers in
Image processing is one of the the last decade have manifest the efficiency
hottest research field and researchers from and accuracy of CBIR based techniques;
the last decagon have proposed many however, CBIR still lacks to understand the
efficient approaches for image analysis such semantic analysis of the image.CBIR system
as [6, 7, 14 ] and retrieval [9, 11, 16]. Image will not be able to map human abstraction
retrieval systems are usually based on into image feature (i.e. creating a semantic
keywords or text data based [4, 18, 19] gap between the low-level image features
where the retrieval is done based on the and high-level human understandable
textual description of the images. The concepts). In this paper, we have proposed
description about the image is usually Semantic Image Retrieval. A semantic based
implemented by the user. Most common approach which uses domain specific for
search engines such as Google and yahoo image retrieval relevant to the user query.
used keyword based research techniques;
this approach is fast and effective. The
keyword based image retrieval system
matches user text query to the textual
2. RELATED WORK performance of most CBIR systems is
forced by these features because they cannot
A lot of experimentation has been efficiently model the expectations of the
conducted on Image Retrieval (IR) on the user. All of existing studies helped in
basis of content similarity. Many techniques improving the results of content based
have been used to enhance the results of images retrieval and lowering down the
image search. These approaches include semantic gap between the user requirements
hierarchical knowledge-based systems for and the search results.
Image Retrieval as experimented by Kurtz,
Camille, et al [40] in 2014. The semantic 3. METHODOLOGY
gap between the low-level image features
and their high level semantics has always The first process of providing the input to
ruined the retrieval quality. So to cope up the SIR is text based. In this approach the
with this problem, Fernandez Miriam et al. user is required to enter the text containing
[36] used an ontology based approach for the information about the thing that user
the enhancements of the image semantics. wants to search. This approach is commonly
This research aimed to solve the restriction used in the current search engines, e.g.
of the keyword based searching to support Google, Bing. The main focus of
the semantic based Image Retrieval. The incorporating this approach in SIR is to
concept of semantic indexing has also been provide ease to the users as they do not have
studied in the field of ontology based to learn the new way of interacting with the
retrieval systems. The literature review on SIR. The user has to simply write down the
Image Retrieval based on semantic concepts text query.
by Riad Alaa et al. [38] had a great Impact The second process of providing the input to
on the Image Retrieval field as it was very the SIR is image based. In this approach the
helpful for improving the semantic image user is required to provide the image of the
retrieval systems accuracy. In this research object(s) which user wants to search. The
various image search techniques are input image can contain a single object or
described for reduction of semantic gap. multiple objects. The user is also provided
Furthermore, based on existing methods and some options (optional) to describe the input
application requirements author have image. This approach is feasible when the
suggested few future assessments. Another user wants to search related objects / images.
important survey was conducted by Liu
Ying et al. [39] in 2007 about the recent
technical achievements on semantic based
Image Retrieval; majority of the recent
publications were included as the test data
for the survey covering diverse amount of
aspects in this area. Similar work has also
been conducted on medical images by Xu J
et al. in [41], the authors focused on the key
features of the image (e.g., shape, texture) in
this research. The authors concluded that the
3.2 Natural language processing

In this paper, we propose a statistical natural


IMAGE DATA language processing model to facilitate
image retrieval, which integrates semantic
information provided by WorldNet and
Low-level visual features. Word Net is an
Preprocessing
online lexical reference system developed
by the Cognitive Science Laboratory at
Natural language processing Princeton University. In its database,
English nouns and verbs are organized into
hierarchies based on the hypernymy
/hyponymy relation between synonym sets,
Jaccard each representing one underlying lexical
concept. A sense is a meaning of a word in
WorldNet, and each sense of a word is in a
K MEANS CLUSTERING different synonym sets. There are one or
several senses of a word, and its synonyms
SEMANTIC IMAGE are ordered by the estimated frequency.
RETRIEVED Barnard et al. used Word Net hyponyms, the
Fig 1: Proposed Methodology IS_A relation, to locate a particular sense of
a word and use it to make the annotation of
3.1 Preprocessing
an image unambiguous. In our system, the
The images revolved by search engine semantic hierarchy of word senses from
may not be relevant to the user query; Word Net is used to strengthen the
therefore the content of each image needs to association between an image and the
be verified. This module is responsible to textual description of a concept. Specifically
check the images for the concurrence with Word Net synonyms are used to trace the
the input query. The objects in each image multiple senses of keywords and establish
are detected using shape based feature their semantic relations. Synonyms are
extraction and these objects are passed to better choice in our model when compared
with hyponyms because they are ordered by
1) Color based Feature Extraction technique the estimated frequency in Word Net and
which uses k-means clustering proposed by duplicated synonyms indicate strong
Guang-Hai Liu et al in to calculate the pixel semantic relations. Semantic measure is
color and color value of the objects and done using NLP.
2) Texture classification technique proposed
3.3 Jaccard
by Mohsen Zand et al in to identify texture /
We will study how to define the
pattern (if any) in the objects.
distance between sets, specifically with the
Jaccard distance. The matching score of two
images is the Jaccard’s coefficient of their
matrices, which is the number of matrix Input to a hierarchical algorithm is an n*n
elements that are equal, divided by the similarity matrix, where n is the number of
objects to be clustered. On the other hand, a
maximum number.
partitional algorithm can use either an n*d
pattern matrix,
Similarity Measures:- Where n objects are embedded in a d-
Our evaluation metric is a generalized dimensional feature space, or an n*n
version of the Jaccard similarity metric of P similarity matrix. Note that a similarity
and Q augmented with the forgiveness rate matrix can be easily derived from a pattern
and with weights on P − Q and Q − P matrix, but ordination methods such as
(misses and false positives, in our case). multi-dimensional scaling (MDS) are
This evaluation formula provides a Texible needed to derive a pattern matrix from a
way to evaluate the multi label classification Similarity matrix. The most well-known
results for both conservative and aggressive hierarchical algorithms are single-link and
tasks. complete-link; the most popular and the
simplest partitional algorithm is K-means.
Since partitional algorithms are preferred in
3.4 k-means clustering:- pattern recognition due to the nature of
Clustering algorithm has been widely used available data, K-means has a rich and
in computer vision such as image diverse history as it was independently
segmentation and database organization. discovered in different scientific fields, it is
The purpose of clustering is to group images one of the most widely used algorithms for
whose feature vectors are similar by clustering.
similarity judgment standard; meanwhile to
separate the dissimilar images. Clustering
algorithms can be broadly divided into two 3.5 semantic images retrieved
groups: hierarchical and partitional.
Hierarchical clustering algorithms In recent years object recognition has
recursively find nested clusters either in advanced significantly. As a direct
agglomerative mode (starting with each data consequence the semantic gap which
point in its own cluster and merging the exists between low-level image features
most similar pair of clusters successively to and high level semantic content of the
form a cluster hierarchy) or in divisive (top- images has been narrowed. The main
down) mode (starting with all the data points idea is to use image classification
in one cluster and recursively dividing each methods to automatically label the image
cluster into smaller clusters). Compared to with semantically relevant labels.
hierarchical clustering algorithms, These methods do not transform the
partitional clustering algorithms find all the histogram into semantic classes. In this
clusters simultaneously as a partition of the section, we shortly describe our approach to
data and do not impose a hierarchical semantic image representation. In particular,
structure. we will discuss in detail how we combined
several cues, in particular shape and color,
into a single image representation.
The parameters of GA used for finding the
results in table 2 are

Comparison of precision, recall, accuracy 1.Probability of crossover Pc=0.7


Similarity. 2.Probability of mutation Pm=0.01
Ever since Jaccard proposed a similarity 3.Number of iterations=150
measure to classify ecological species in
1901, numerous binary similarity and Graphical comparison is shown in figure 1.
distance measures have been proposed in
various fields.

Table 1. The three similarity coefficients

Similarity Coefficient Actual Formula


(X,Y)

1
Recall recallC = |DC | x∈DC

Fig 1 Comparison of similarity


coefficients for fitness value
1 We are experimented the different
precisionC = DC D score(Px);
Precision | |
x

∈C
coefficient of precision, recall, accuracy.
The results are defined under the actual
formula.
4 Result

Initially for the experimentation, we


1 trained Semantic Image Retrieval (SIR) and
accuracyD = |D | score(Px):
x∈D built the concepts using 900 images which
Accuracy
contain pictures of 20 different mammals.
Partial training dataset is shown in figure 5.
We have evaluated SIR on large number of
test cases; results were promising and
showed the efficiency of the proposed
system. In this section, few of the test cases
are presented and discussed in detail.
To evaluate recall and precision of each base class,
we extend the classic definitions.
proposed approach with Reference to four
different test cases. As depicted our
proposed hybrid approach outperforms these
approaches with reference to accuracy.
the percentage improvement of proposed
hybrid technique over number of test cases;
the proposed solution improvement
percentage varies over number of test cases;
this is because the content of images present
in each test case plays an important role.

Partial Training set

False Positive Percentage vs. Test Cases

The false positive percentage over number


NLP AND JACCARD USING of test cases, the proposed solution false
SEMANTIC IMAGE RETRIEVAL positive percentage ranges from 0.60 to 2
percent in the test cases which shows the
result accuracy of the proposed solution.
5 CONCLUSION

Image retrieval systems are usually


based on keywords or text meta data based.
Most common search engines such as
Google and Bing are based on keyword
based search techniques. This approach is
fast and effective; however it still has some
disadvantages. Content based Image
retrieval (CBIR) has been studied for many
years which focuses on extracting and
Percentage Improvement vs. Test Cases comparing features from the images.
Researchers in the last decade have
The accuracy comparison of color demonstrated the efficiency and accuracy of
based shape based, texture based and our CBIR based techniques, however, CBIR still
lacks to understand the semantic analysis of [6] Xu J, Faruque J, Beaulieu CF, Rubin DL,
the image.NLP extracts the keyword from Napel S. A comprehensive descriptor of
user given query text. Based on extracted shape: method and application to content-
keyword, similarity is measured by jaccard based retrieval of similar appearing lesions
similarity coefficient measure.. The k-means in medical images. J Digit Imaging 2012;
clustering algorithm to image Retrieval 25:121–8.
system Classification is done by modified
RBF which performs classification based on [7] van de Weijer, J., Schmid, C.: Applying
kernel point. Scenic name, label words, color names to image description. In: IEEE
image are used as reasoning condition for International Conference on Image
image retrieval. The proposed system has Processing (ICIP), San Antonio, USA
been tested on large number of test cases; (2007).
experimental results show the efficiency and
effectiveness of the proposed technique.
[8] Felzenszwalb, P.F., McAllester, D.A.,
REFERENCES Ramanan, D.: A discriminatively trained,
multiscale, deformable part model. In: IEEE
[1] Shotton J, Blake A, Chipolla R, Computer Vision and Pattern Recognition.
Contour-based learning for object detection. (2008).
In Proceeding of International Conference
on Computer Vision (2005). [9] Lowe, D.: Distinctive image features
from scale-invariant key points.
[2]. Fellbaum, “WordNet: An Electronic International Journal of Computer Vision
Lexical Database,”MIT Press, 1998. (IJCV) 60 (2004) 91–110.

[10] Gevers, T., Smeulders, A.: Color based


[3] . Dacheng Tao, Dianhui Wang, Fionn object recognition. Pattern Recognition 32
Murtagh, "Machine learning in intelligent
(1999)453–464.
image processing", Signal Processing,
Volume 93, Issue 6, June 2013, Pages 1399-
1400.

[4].K. Barnard, P. Duygulu, and D.


Forsyth, “Clustering Art,” Computer
Vision and Pattern Recognition, pp. II:
434-439, 2001.

[5] Mihalcea, R. and Moldovan, D.:


Semantic indexing using WordNet senses. In
Proceedings of ACL Workshop on IR
& NLP, Hong Kong, October 2000.

Вам также может понравиться