International Journal of Recent Technology and Engineering (IJRTE)

ISSN: 2277-3878, Volume-8 Issue-6, March 2020

Intelligent Water Drop Algorithm based

Relevant Image Fetching using Histogram and
Annotation Features
Saket Jain, Rajendra Gupta

 [3],[4],[5],[6],[7]. Though, there is a space between the

Abstract: Social media network increase trend of image semantic interest of the customer and the mined visual
collection at various platforms. Hence getting an relevant image characteristic signifiers. For example, as shown in Figure 1, if
as per query image or text is depend on retrieval algorithms. the customer gives an image having a red apple as an inquiry,
Number of researcher has proposed algorithms for fetching
the retrieved pictures may include red flowers, red balloon or
relevant images, but relevancy of those still need improvement.
Hence proposed paper has utilized the Intelligent water Drop green fruit depending on the visual descriptor utilized by the
algorithm for initial clustering of images as per feature values. CBIR. An additional inquiry in particular image retrieval
Clustering or relevancy of an image depends on visual feature comes out when the inquiry image have a number of objects.
histogram and annotation similarity. Here property of moving a In fact, the recovered images may not be related to the exact
water drop from one node to another in a water drop graph has objective meant by the customer.
increase the clustering accuracy of the work. Experiment was
Lately, the several query retrieval systems has been
done on real dataset having five different group of image set with
annotation. Result shows that proposed work has increase the projected in order to fill up this semantic gap, and improve
retrieval relevancy accuracy as well as reduce the fetching of the the retrieval presentation [8]. For multiple query retrieval
images. This reduction of time was obtained by using the system, the customer expresses his attention via a collection
clustering structure of image dataset. of query images. This gives a more affluent perceptive of the
customer high-level interest to the retrieval system and fills
Keywords: Digital Image Processing, Information Extraction, out the semantic gap with the low-level image function.
feature extraction, Re-ranking.
Dissimilar to solo query based CBIR system where the gaps
between the visual attribute descriptor of the inquiry image
I. INTRODUCTION
and the visual characteristic descriptor of every picture in the
In the last few years, the ground of computer vision (CV) record are calculated and arranged in order to offer the nearly
and Machine learning (ML) has widen its scale to address a all related images to the customer, for multiple query image
variety of duties in Textile and Fashion manufacturing based CBIR systems, the dispute is how to calculate the gap
industry [1, 2]. An illustration of such type of classical duty is between the collection of inquiry images and every image in
to take out visually related pictures from a huge picture the record in a way that improves the retrieval end results and
record in reply to a question image given by customer. The reproduce the elevated semantic.
increase of visual search brings forwards a variety of search The remaining research paper is prepared as follows.
engines TinEye, Flicker, Google as well as multimedia study Section II explains color form. Section III shows histogram
paradigms [3,4]. and picture retrieval. Section IV and Section V give aspects
of color quantization and image query, equally. Section VI
This huge boost triggered the challenge of mining highlights trial results and finally, Section VII illustrates the
particular image amongst vast groups. Thus, image retrieval generally end result of this document.
has become a vigorous ground of research [1]. Content Based
Image Retrieval (CBIR) permits the consumers to convey his II. RELATED WORK
field interest by inserting a query picture that reproduce the
Bhute and meshram (2013) [9] designed a new content
semantics he/she is searching for. The record is then retrieved
based picture indexing and revival structure. In this
according to the inquiry content. For CBIR systems, the
occupation, utilize color, texture and form of inputs to
image properties of a picture are explained utilizing low-level
facilitate the mending process. For improved mining, extract
characteristic signifier [2]. More particularly, the signifiers
the color, texture and character information of input data
can be the shade, the quality and the shape types of the
normally using boundary detection which is widely utilized
picture. These low-level characteristics interpret the visible
in sign processing and image firmness. In color characteristic
content of the picture into statistical vectors that permit
removal scheme, image histogram and color correlogram are
quantitative evaluation of the likeness between both pictures
fast retrieval performs the antipole-tree system for indexing
the descriptions. Finally histogram Euclidean division is
Revised Manuscript Received on February 01, 2020. executed to figure out the similarity amongst record and
* Correspondence Author
Saket Jain *, Dept. of Computer Science & Engg., Rabindranath Tagore
inquiry image.
University, Bhopal, India. Email: [email protected]
Dr. Rajendra Gupta, Dept. of Computer Science & Engg, Rabindranath
Tagore University, Bhopal, India. Email: [email protected]

Retrieval Number: F6983038620/2020©BEIESP Published By:

DOI:10.35940/ijrte.F6983.038620 Blue Eyes Intelligence Engineering
23 & Sciences Publication
 Intelligent Water Drop Algorithm based Relevant Image Fetching using Histogram and Annotation
Features

Munje and kapgate (2014) [10] planned a new CBIR clothing. Their dataset is collected of 684 pictures of sizes
system. In CBIR system, the color and texture information’s that series between 480x360 and 1280x720 pixels gathered
are extracted from inquiry images. The color characteristic is from 15 videos of YouTube. According to them R-CNN
extracted by utilizing color instant and surface features are gained highest accurateness of around 85%. Job has also been
extracted by using gabor sort out system and wavelet convert. done for detection of diverse form of dresses.
The information’s are calculated for all pictures of inquiry
and record images. Completely 15 characteristics for each III. PROPOSED METHODOLOGY
image are determined, 6 color characteristic and enduring are
This section gives a complete explanation of proposed
texture element. Lastly according to the resemblance
IRIWD (Image Retrieval By Intelligent Water Drop). Here
calculation the important images are taken out.
fig. 1 shows steps of developing a ontology from the image
In [11] uses the visual contents of an image similar to
database. Image set of features were extract and store in
worldwide features-color characteristic, shape feature,
hierarchal structure where selection of cluster center were
texture feature, and local features-spatial domain present to
done by Intelligent Water Drop Algorithm. Two features
signify and index the image. CBIR method combines global
were extract for developing of ontology from the image
and local features. In this paper worked on Haar Discrete
dataset first was annotation second was histogram. In other
Wavelet Transform (HDWT) for decaying an image into
part of this section testing was perform where developed
horizontal, vertical and diagonal region and Gray Level Co
ontology gives an output as per input image and test query.
occurrence Matrix (GLCM) for feature extraction. Support
Vector Machine (SVM) used, diverse calculations to recover A. Visual Content Processing
the accuracy and implementation of recovery. Input dataset may have different dimension image
In [12] pictures resize according to the section of interest collection so transformation of all set of images in same
for the earlier recovery of pictures. Removing and Row-X-Column is prior requirement. As this work extract
eliminating difficult background will boost up further image image visual features so all image matrix are of equal size.
processing. Very well-built discriminative power For annotation feature text pre-processing steps were apply
characteristic makes an important element in image and such as conversion of string to words and then removal of
video recovery. Consequently, it is extremely significant to stop words form the annotation.
discover an effectual technique to calculate the directionality
of an image, and tamura utilizes statistical calculation to B. Feature Extraction
compute statistical characteristic. And thus we mine texture In this work two type of features were used for the image
characteristics and shape and fused these element vectors of retrieval first was visual where histogram values were
tamura and form combinations for enhanced end result obtained. Here work has utilized B bins of histogram values.
Fu et al (2016) [13] projected the Convolution Neural So image feature is counting of pixels range in [(1-B), (B+1 –
Network (CNN) based deep characteristic mining techniques. 2B), ………(PB-M)], where M is max pixel limit and P is
The CBIR structure utilizes the direct Support Vector (M/B – 1). This can be understand as let image have 256 type
Machine (SVM) to arrange a hyper plane which can separate of pixel values, now bins have values in range of [(0-15),
the relative images sets and dissimilar images sets to a huge (16-31), (32-47),………………….(250-255)]. Small feature
degree. The couple of features from key image and each set of sixteen values were produced with the image in form of
study image in the image dataset are known as information visual content so comparison takes less time.
(input). The analysis images at that end are evaluated by the
division amongst the couple features and the specialized
hyper plane. Tests reveals that the intended system can
significantly boost the implementation of CBIR for image
revival undertakings.
Alsmadi (2017) [14] intended a novel resemblance system
for CBIR utilizing mimetic approach. In this work, color,
character and color texture information are extracted from
question images. The shape characteristics are utilized to
mine the belongings of the shape of the images. The texture
features are mined by utilizing GLCM which is vigorous
image statistical investigation approach. Then the likeness is
calculated amongst the mined function and record feature by
utilizing mimetic algorithm. Lastly the presentation of the
work is examined.
In [15] writer attempted texture based image retrieval
(TBIR) utilizing machine learning algorithms and their
amalgamation including Faster Region based Convolution
Neural Network (R-CNN), Adaptive Linear Binary Pattern
(ALBP), Complete Local Oriented Statistical Information
Booster (CLOSIB) Histogram of Oriented Gradients (HOG)
and Half Complete Local Oriented Statistical Information
Booster (HCLOSIB) for neighboring patch explanation of

Published By:
Retrieval Number: F6983038620/2020©BEIESP
Blue Eyes Intelligence Engineering
DOI:10.35940/ijrte.F6983.038620
& Sciences Publication
24
 International Journal of Recent Technology and Engineering (IJRTE)
ISSN: 2277-3878, Volume-8 Issue-6, March 2020

accuracy. Hence feature set is collection of two type of data

first is bins count and second was words from annotation.
Image Dataset Feature [Bins Count, Annotations]

A. IRIWD (Image Retrieval Intelligent Water Drop)

In this algorithm each image act as ‘WATER DROP’
where graph was develop having drop as node and feature
difference as edge. This edge in the IWD algorithm s term as
Pre-Processing
‘SOIL’. So a drop move form one node to other such that soil
between nodes should be less as soil reduce the velocity of
the moving drop. Hence as per similarity between the drops
soil is less and cluster get form.
B. Water Drop Graph
Visual Feature Annotation This is complete graph having N number of nodes where N
is total number of images in the dataset. Hence WDG is NxN.
Graph weight value in form of SOIL S was estimate for each
edge between nodes. This value was obtain by the normalized
value of annotation similarity and histogram distance of
Develop Water between nodes (Images).
Drop Graph
--Eq.
Generate 1
Population Where are normalization factor whose value range in 0.1 to
0.0001 as distance value by Euclidian is high so
normalization is required for annotation factor.
C. Static and Dynamic Parameter
In this step some of constant were initialize before the start
T
iteration of algorithm such as soil updating parameters S1 = 1, S2 = .01,
and S3 = 1, velocity updating parameters V1 = 1, V2 = .01, and
V3 = 1. Finally global and local soil constants βL and βG are
initialize by 0.9. Values of constants may be vary as per
Selection algorithm requirement.
Probability D. Generate population
As population is collection of chromosomes where each
chromosome is collection of cluster centers. Hence
chromosome having set of image features which act as cluster
Update Drop center. To better understand this let image dataset have N
Velocity number of images so one of possible solution is Chromo =
{I1, I2, I9, IM} where M is class of images. In similar way
other set of M images were collect from dataset randomly.
Update Drop Soil Hence population generation function can be written as:
E. PGenerate_Population(D, M, N)
F. Selection Probability
This is probability of the drop to move from one of N-1
Fitness Value possible nodes. So this probability value SP is calculate as
below:

Crossover

Cluster Dataset

Fig.1. Block diagram of proposed model. G. Update Velocity

Further each annotation string is convert into set of words as Update velocity of the ith drop moving toward jth node by
per text pre-processing, so similar kind of image may have below formula:
different set of words. This annotation increase searching

Retrieval Number: F6983038620/2020©BEIESP Published By:

DOI:10.35940/ijrte.F6983.038620 Blue Eyes Intelligence Engineering
25 & Sciences Publication
 Intelligent Water Drop Algorithm based Relevant Image Fetching using Histogram and Annotation
Features

L. Cluster Dataset
After T number of genetic algorithm iteration final update
H. Update Soil population was obtain. Best fitness value chromosome gives
an set of cluster centers. Here all other images were clustered
Update velocity of the ith drop moving toward jth node by
accordingly as per cluster center feature values.
below formula:
M. Testing Phase
Once dataset get grouped into cluster form than testing
dataset will pass and evaluate the resultant ranked images. So
each image from the testing dataset is pre-process first as
HD is heuristic durability a constant value range in 0-1. done in learning phase, further similar visual, annotation
features were also extract. Finally based on testing image
feature cluster center feature values were compared and most
I. Fitness Function matching cluster is select for the image ranking. Now each
For finding the fitness value of the chromosome one need clustered image features were compared with testing image
to compare the cluster fitness value. As work uses visual and feature for final rank of images. This comparison was done
annotation feature of the image so fitness value takes of them by fitness function.
as input for each chromosome fitness value evaluation. N. Proposed Algorithm
Algorithm 1 shows the evaluation steps of fitness value from Input: D // Dataset
set of dataset images, as per population Output: CD // Clustered Dataset
1. Loop 1:N
Input: P, D // P: Pupulation, D: Dataset of N images 2. DImage-Pre-Processing(D[n])
Output: Fitness 3. DText-Pre-Processing(D[n])
Loop 1:P 4. FFeature-Extraction(D[N])
Loop 1:N 5. End Loop
F[N,P]Distance(Chromo[P], Feature[N]) 6. WDGGraph(N, F)
EndLoop 7. Initialize Static and Dynamic parameters
MinMinimum(F[N,:]) 8. PGenerate-Population(D)
Fitness[P]Sum(Min) 9. Loop 1:T // T: Number of iteration
EndLoop 10. SPSelection Probability(WDG)
11. Loop 1:M
Where Distance is function that find difference in Bins count 12. WDGUpdate_Velocity(SP, WDG)
value of two images (Cluster center image, Input Image), 13. WDGUpdate_Soil(SP, WDG)
while dissimilar words from the annotation were also 14. End Loop
increase this distance value. But as distance value of both 15. FitnessFitness-Function(P)
feature is quit high so normalization of feature values were 16. GBest(Fitness)// G: Global
done by multiplying the Bins value with an constant range in 17. PCrossover(G, P)
18. End Loop
[0.01 to 0.0001].
19. FitnessFitness-Function(P)
J. Global Crossover 20. GBest(Fitness)// G: Global
In this step of genetic algorithm crossover of the algorithm 21. CDCluster(G, D)
was done by selecting one common parent in all crossover
with other set of chromosome. So selection of this common IV. EXPERIMENT AND RESULT
parent depends on fitness value. Here best fitness values A. Experimental Setup
chromosome act as common parent in all crossover Experiment was done on MATLAB platform where
operation. So other set of chromosome undergoes crossover machine have configuration of 8GB Random Access
by randomly replacing a cluster center as per common parent memory, i5 processor. Results were evaluate on real dataset
cluster center value in same position. So if best set of Chromo having 100 Images from 5 category [14]. Each category have
is {I1, I2, I9, I22} and random position is two than I2 is place in 20 image in a set.
same position two of other parent chromosome. This Detail description of dataset is shown in table 2.
replacement is done only if other parent do not have same Feature Description
cluster center in other positions of chromosome. Number of Images 100
Category 5
K. Population Updation Dimension 384x256
As crossover changes the chromosomes of the population Dimension Three Dimension Color
so retention of this chromosome depends on fitness value.
This can be understand if child chromosome have good
fitness value as compared to parent fitness value than new
child was include in the population, otherwise parent
chromosome will continue in population. Hence in all
situation population size will never change from P number.

Published By:
Retrieval Number: F6983038620/2020©BEIESP
Blue Eyes Intelligence Engineering
DOI:10.35940/ijrte.F6983.038620
& Sciences Publication
26
 International Journal of Recent Technology and Engineering (IJRTE)
ISSN: 2277-3878, Volume-8 Issue-6, March 2020

Table –III: Sample Image with Category. and annotation.

Category Sample Image D. Results
Table –IV: Precision Value Based Comparison of Proposed
Human Model for top 18 images.
Category Previous Work GA IRIWD

Human
0.5 0.861117 0.944444
Building
0.362856 0.525 0.584418
Transport
0.41667 0.58335 0.77778
Animal
0.447 0.8333 1
Building Food 1
0.5278 1

Above table 4 represent that accuracy of relevant image

retrieval of proposed IRIWD model using IWD genetic
algorithm was high as compared to previous approach. Here
use of visual and annotation feature in cluster form increase
this efficiency of work. In previous work [13] Harris feature
Transport reduce retrieval accuracy.

Table –V: NDCG@18 Value Based Comparison of Proposed

Model.
Category Previous Work GA IRIWD

Human
0.64646 0.898792 0.958853
Building
0.460468 0.60683 0.695542
Transport
Animal 0.576752 0.633327 0.787896
Animal
0.62359 0.889843 1
Food
0.668422

Above table 5 represent that NDCG value for top 18 resultant

images and it was obtained that proposed IRIWD model
using IWD genetic algorithm was high as compared to
previous approach. Here use of visual and annotation feature
in cluster form increase this relevancy of resultant ranked
Food images. In previous work [13] corner feature point matching
reduce this NDCG value.

Table –VI: Precision Value Based Comparison of Proposed

Model for top 20.
Category Previous Work GA IRIWD

Human
0.475 0.85 0.925
Building
0.275 0.525 0.575
Transport
A. Evaluation Parameters 0.375 0.575 0.775
Normalized Discounted Cumulative Gain (NDCG): l is Animal
0.425 0.75 1
list of relevant and ir-relevant vector having 1/0 values for ith
Food 1
position. i range from 1 to P. 0.475 0.9

Above table 6 represent that accuracy of relevant image

retrieval of proposed IRIWD model using IWD genetic
B. Precision: In this parameter all relevant images are algorithm was high as compared to previous approach. Here
divide by total number of ranked P images. use of visual and annotation feature in cluster form increase
P = Sum(l) / P this efficiency of work.
C. Execution Time (Seconds): This is time required to
fetch image from the dataset as per input testing image

Retrieval Number: F6983038620/2020©BEIESP Published By:

DOI:10.35940/ijrte.F6983.038620 Blue Eyes Intelligence Engineering
27 & Sciences Publication
 Intelligent Water Drop Algorithm based Relevant Image Fetching using Histogram and Annotation
Features

In previous work [13] harris feature reduce retrieval

accuracy. Table –IX: Average Execution Time in Seconds Based
Comparison of Proposed Model.
Table –VII: NDCG@20 Value Based Comparison of Category Previous Work GA IRIWD
Proposed Model.
Category Previous Work GA IRIWD Human 0.898 0.7781 0.7339

Human Building 0.8923 0.7922 0.7399

0.620478 0.889222 0.945367
Building Transport 0.8103 0.7613 0.7269
0.433405 0.600128 0.666623
Transport Animal 0.8383 0.7838 0.7439
0.539146 0.624897 0.785294
Animal Food 0.8987 0.7938 0.7539
0.58293 0.831823
Food
0.624838 0.934797 Above table 9 represent that execution time of relevant
image retrieval of proposed IRWD model using IWD genetic
Above table 7 represent that NDCG value for top 18 algorithm was low as compared to previous approach. Here
resultant images and it was obtained that proposed IRIWD use of cluster structure for image fetching reduce this time.
model using IWD genetic algorithm was high as compared to
previous approach. Here use of visual and annotation feature V. CONCLUSIONS
in cluster form increase this relevancy of resultant ranked
images. In previous work [13] corner feature point matching In the research of Image retrieval, there are a lot of
achievements in image semantic feature, they can be applied
reduce this NDCG value.
to content-based image retrieval to analyze the transition
between visual features and semantic features of the images.
Table –VIII: Average Precision Value Based
This paper has proposed a clustered structure to reduce the
Comparison of Proposed Model.
Category Previous Work GA IRIWD execution time of the algorithm for retrieval of relevant
images. Here visual content histogram and annotation
Human features were used by the proposed IRIWD model for finding
0.4875 0.855558 0.934722
the fitness value during clustering of images. Experiment was
Building
0.318928 0.525 0.579709 done on real dataset having different category images.
Transport Results were compared with previous existing approaches
0.395835 0.579175 0.77639 and it was obtained that proposed method has improved the
Animal
0.436 0.79165 1
NDCG value by 10.51%, while accuracy improved by
Food 13.73%. In future one can involve other visual feature to
0.5014 0.95 1 increase this accuracy of work.

Above table 8 shows average values for different top REFERENCES

relevant image set, proposed IRIWD model using genetic 1. Ben Ismail M. M. "A Survey on Content-based Image Retrieval."
algorithm was high as compared to previous approach. Here International Journal of Advanced Computer Science and Applications,
use of visual and annotation feature in cluster form increase Vol. 8, No. 5, pp. 159-170, 2017. [2] J. Komal , V. Akhilesh, G. Savita,
Swati Goel. "A Survey on Recent Image Indexing and Retrieval
this efficiency of work.
Techniques for Low-Level Feature Extraction in CBIR Systems," in
IEEE International Conference on Computational Intelligence &
Communication Technology (CICT), 2015.
2. Julie Verma, Urooj Fatima, ” A Review on Research and Challenges in
Modern Image Retrieval Techniques in Big Data Systems” ,
InternationalJournal ofResearch and Development in Applied Sciences
and Engineering (IJRDASE) ISSN: 2454-6844 Volume 13, Issue 1,
March 2017.
3. Mohamed M.BenIsmail ,"A Survey on Content-based Image Retrieval”,
(IJACSA) International Journal of Advanced Computer Science and
Applications, Vol.8, No.5, 2017.
4. M. Eitz, K. Hildebrand, T. Boubekeur et al., “Sketch-Based Image
Retrieval: Benchmark and Bag-of-Features Descriptors,” IEEE
Transactions on Visualization & Computer Graphics, vol. 17, no. 11, pp.
1624-1636, 2011.
5. [25] G. H. Liu, Z. Y. Li, L. Zhang et al., “Image Retrieval Based on
Micro-structure Descriptor,” Pattern Recognition, vol. 44, no. 9, pp.
2123-2133, 2011.
6. [26] G. H. Liu, L. Zhang, Y. K. Hou et al., “Image Retrieval Based on
Multi-texton Histogram,” Pattern Recognition, vol. 43, no. 7, pp.
2380-2389, 2010.

Fig. 2. Average NDCG value Based Comparison.

Published By:
Retrieval Number: F6983038620/2020©BEIESP
Blue Eyes Intelligence Engineering
DOI:10.35940/ijrte.F6983.038620
& Sciences Publication
28
 International Journal of Recent Technology and Engineering (IJRTE)
ISSN: 2277-3878, Volume-8 Issue-6, March 2020

7. H. Wei, Y. Li, W. Y. Chau et al., “Trademark Image Retrieval Using

Synthetic Features for Describing Global Shape and Interior Structure,”
Pattern Recognition, vol. 42, no. 3, pp. 386-394, 2010.
8. L. Wu, R. Jin, and A. Jain, “Tag Completion for Image Retrieval,” IEEE
Transactions on Pattern Analysis & Machine Intelligence, vol. 35, no. 3,
pp. 716-727, 2013.
9. Avinash N Bhute1, B. B. Meshram,”Content Based Image Indexing and
Retrieval”International Journal of Graphics & Image Processing |Vol
3|issue 4|Nov. 2013
10. Munje, P.N. and KAPGATE, D., 2014. novel content– based image
retrieval techniques for medical database. International Journal, 3(6).
11. Ekta Gupta, Rajendra Singh Kushwah “Combination of Global and
Local Features using DWT with SVM for CBIR”. PAPER:
978-1-4673-7231-2/15/$31.00 ©2015 IEEE
12. Neelima Bagri and Punit Kumar Johari,” A Comparative Study on
Feature Extraction using Texture and Shape for Content Based Image
Retrieval”, Vol.80 (2015), pp.41- 52,ijast,2015.
13. Fu, R., Li, B., Gao, Y. and Wang, P., 2016, October. Content-based
image retrieval based on CNN and SVM. In Computer and
Communications (ICCC), 2016 2nd IEEE International Conference on
(pp. 638-642). IEEE.
14. Alsmadi , “An efficient similarity measure for content based image
retrieval using memetic algorithm”, Egyptian Journal of Basic and
Applied Sciences 4 (2017) pp.112–122.
15. Olalla, O.G.; Alegre, E.; Robles, L.F.; Fidalgo, E.; Saikia, S. Textile
Retrieval Based on Image Content from CDC and Webcam Cameras in
Indoor Environments, Sensors, 2018; doi:10.3390/s18051329
16. Jiaohua Qin, Hao Li, Xuyu Xiang, Yun Tan, Wenyan Pan, Wentao Ma1,
And Neal N. Xiong. “An Encrypted Image Retrieval Method Based On
Harris Corner Optimization And Lsh In Cloud Computing”. Ieee Access
March 7, 2019.
17. http://wang.ist.psu.edu/docs/related.shtml

Retrieval Number: F6983038620/2020©BEIESP Published By:

DOI:10.35940/ijrte.F6983.038620 Blue Eyes Intelligence Engineering
29 & Sciences Publication