See discussions, stats, and author profiles for this publication at: [Link]

net/publication/310503949

Real-time recognition of cattle using animal biometrics

Article in Journal of Real-Time Image Processing · September 2017

DOI: 10.1007/s11554-016-0645-4

CITATIONS READS

63 2,165

5 authors, including:

Santosh Kumar Sanjay Kumar Singh

IIIT Naya Raipur, Chhattisgarh Indian Institute of Technology BHU
153 PUBLICATIONS 1,827 CITATIONS 194 PUBLICATIONS 4,450 CITATIONS

SEE PROFILE SEE PROFILE

Ravi Shankar Singh Amit Kumar Singh

Shri Venkateshwara University National Institute of Technology Patna
87 PUBLICATIONS 844 CITATIONS 199 PUBLICATIONS 6,993 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Santosh Kumar on 24 June 2020.

The user has requested enhancement of the downloaded file.

Real-time recognition of cattle using animal
biometrics

Santosh Kumar, Sanjay Kumar Singh,

Ravi Shankar Singh, Amit Kumar Singh
& Shrikant Tiwari

Journal of Real-Time Image

Processing

ISSN 1861-8200

J Real-Time Image Proc

DOI 10.1007/s11554-016-0645-4

1 23
Your article is protected by copyright and
all rights are held exclusively by Springer-
Verlag Berlin Heidelberg. This e-offprint is
for personal use only and shall not be self-
archived in electronic repositories. If you wish
to self-archive your article, please use the
accepted manuscript version for posting on
your own website. You may further deposit
the accepted manuscript version in any
repository, provided it is only made publicly
available 12 months after official publication
or later and provided acknowledgement is
given to the original source of publication
and a link is inserted to the published article
on Springer's website. The link must be
accompanied by the following text: "The final
publication is available at [Link]”.

1 23
 Author's personal copy
J Real-Time Image Proc
DOI 10.1007/s11554-016-0645-4

SPECIAL ISSUE PAPER

Real-time recognition of cattle using animal biometrics

Santosh Kumar1 • Sanjay Kumar Singh1 • Ravi Shankar Singh1 •

Amit Kumar Singh2 • Shrikant Tiwari3

Received: 14 June 2016 / Accepted: 3 October 2016

Ó Springer-Verlag Berlin Heidelberg 2016

Abstract With the advent of efficient recognition tech- protocol. The efficacy of proposed muzzle point recognition
niques, animal biometric systems have gained more prolif- approach for cattle evaluates under identification settings
eration for the identification and monitoring of cattle. A and yields 96:87 % recognition accuracy for identifying
cattle biometric system is a pattern recognition-based system individual cattle. The proposed approach also valued the
for the identification of livestock. In this paper, we propose a 10.25 sec recognition time for enrollment and identified
novel muzzle point recognition based on Fisher locality individual cattle on different sizes of muzzle point images.
preserving projection algorithm for the recognition of cattle
in real time. We have captured images of animals using a Keywords Real-time image processing  Animal
surveillance camera and transferred them to the server by biometrics  Muzzle point  Cattle recognition  FLPP 
wireless network technology. The major contributions are as FLDA  SVM  Feature extraction  Classification
follows: (1) preparation of muzzle point database, (2)
extraction of the salient set of features using proposed
muzzle point recognition approach, and (3) evaluation and 1 Introduction
comparison analysis of the introduced method and several
existing recognition algorithms on a standard benchmark Animal recognition systems have received much attention
and proliferation due to the wide range of applications and
& Santosh Kumar
used in the field of animal biometrics, computer vision,
[Link].cse12@[Link] pattern recognition, and cognitive science [34]. The cattle
Sanjay Kumar Singh
recognition system can apply to the variety of applications,
[Link]@[Link] such as animal registration, traceability, monitoring, iden-
Ravi Shankar Singh
tification of missed, swapped of cattle, reallocation of
[Link]@[Link] livestock, and verification of false insurance claims
Amit Kumar Singh
[14, 35]. However, classical animal recognition-based
amit_245singh@[Link] methodologies include the manual approaches, such as ear-
Shrikant Tiwari
tagging, ear tips or notches, freeze-branding, and embed-
shrikanttiwari15@[Link] ded RFID-microchip to recognize the individual cattle [3].
On the other hand, automatic recognition and behavior
1
Department of Computer Science and Engineering, Indian analysis of animal can be done by captured images of ani-
Institute of Technology (Banaras Hindu University),
Varanasi 221005, India
mals [11]. Due to the enormous amount of man-power
2
requirements, classical animal recognition approaches have
Department of Computer Science and Engineering, Jaypee
high cost and venerability loses due to duplication, fraudu-
University of Information Technology, Solan,
Himachal Pradesh 173234, India lent, and forged of embedded standard tags. The human
3 interpretation is subjective in prestigious animal recognition
Department of Computer Science and Engineering, Shri
Shankaracharya Technical Campus, Junwani, approaches [3]. The registration process of the animal pro-
Bhilai, District-Durg, Chattisgarh 490020, India vides techniques to control the various kinds of efforts for

123
 Author's personal copy
J Real-Time Image Proc

the manipulation and swapping of cattle. The traceability problems of cattle using GPS positioning, Bluetooth or Wi-
process provides a safer food supply. It also provides the Fi networks, and RFID Huhtala et al. [25]. However, it
excellent platform for producers and consumers [32]. reported major problems during the practical implementa-
Moreover, traceability process also caters a procedure to tion for embedding the sensors in the animal’s body. The
verify the owner or parentage of animals. It also controls and emerging technologies are unable to provide better security
outbreaks the critical diseases of animals for their health to livestock or cattle. Moreover, these techniques are
management in the livestock [63]. invasive and have the significant problems for embedding
The registration of cattle is one of the essential com- the electronic devices in their body [56].
ponents in the emerging animal biometrics, livestock Besides that, all categories of classical animal recogni-
framework-based systems, and classical animal recogni- tion approach, the artificial marking-based techniques can
tion-based approaches. be duplicated, fraudulent, and forged the embedded a
The classical animal recognition methodologies are ear unique identification number of ear-tags. Therefore, it is
tags, freeze-branding, ear tattoos, ear tips or notches, and unable to recognize and verify the false insurance claims.
embedding of microchips in the animal body. These are Due to the significant limitations and failures of classical
invasive approaches for the recognition of individual cattle. animal recognition-based approaches, there is the need to
However, classical animal recognition-based approaches are design and develop a real-time cattle recognition system
more susceptible to massive vulnerability to loss and illeg- for identification, verification, and monitoring of cattle,
ibility. The classical animal recognition techniques always individual animals and other livestock.
lead more security issues for the protection of cattle or other
animals [29]. The real-time cattle recognition system pro- 1.1 Motivation of work
vides an efficient way to stop the cattle manipulations. In the
classical animal recognition systems, probability of regis- The identification of animals is the major problem for verifi-
tration and identification of animal using ear tags is more. cation of false insurance claims, animal tracking, animal
However, ear tags became damaged, lost, or stolen easily. monitoring, identification of locomotion and posture behavior
The duplication and forge of the labeled unique number in of livestock animal or cattle before calving in the livestock
the ear tags done easily. Therefore, it has more vulnerability framework, and prestigious animal recognition-based systems.
due to losses and major security issue-related ear-tagging In our acknowledgment, there is no such animal bio-
system for identification of different cattle. metrics-based cattle recognition systems available in the
In the classical animal recognition systems, there is public domain. To provide better solutions for identifica-
more probability of registration and recognition of animal tion and verification of false insurance claims, monitoring
using ear tags. However, ear tags became damaged, lost, or of livestock and assistance during health management of
stolen easily. The unique number in the ear tags can be animals, and efficient recognition are required for the
duplicated, forged. It has more vulnerability due to losses prevention of critical diseases, food, feed, and distribution
and major security issue-related ear-tagging system. of cattle in the livestock framework.
In case of false insurance claim verification of animals, These are the major problems of identification and
there is no such real-time animal biometric-based recog- monitoring of animal in the classical animal recognition
nition systems that are available in the literature or public approaches and traditional livestock framework based
domain to cater a better protection, verification, and systems. These problems can not be ignored by various
recognition of individual cattle or livestock (animal) using scientists, veterinary professionals, animal experts, and
their primary animal biometric characteristics. different research communities to contribute valuable
In current state-of-the-art-based animal recognition efforts for the design and development of robust, nonin-
approaches, the animals are verified and identified to solve vasive, and real-time animal biometric-based recognition
the biggest problems of the false insurance claim by cutting systems for identifying individual cattle. Therefore, it is
their ear or snatching the labeled ear tags or notches from required to develop a real-time cattle recognition system
the animals ear by different governmental organizations for identifying and monitoring of different animals.
and private animal insurance providers. It is evident that a In this paper, we propose a muzzle point recognition-based
falsification in the labeled ear tags can make duplication, cattle recognition system for identifying individual animals
forgery, and fraudulent process. Therefore, it is tough to (especially cattle) in the real-time. The images of cattle are
verify and recognize the registered insurance animals captured using surveillance camera. The surveillance cameras
(owner of cattle) or impostor (non-insurance) animal. are deployed in the stables of cattle. The total number of cattle
In classical animal recognition-based systems, various is 500 in the stable of the Department of Dairy and Husbandry,
emerging technologies have been investigated to provide a Institute of Agriculture Science (I.A.S), Bananas Hindu
better solution to identification, monitoring, and tracking University (B.H.U), Varanasi-221005.

123
 Author's personal copy
J Real-Time Image Proc

The proposed cattle recognition system provides better We use wireless networks (i.e., Internet technology) to
solution to recognition and monitoring of cattle or livestock transfer the image of animals (cows) to the server of the
animals. These recognition of cattle using classical animal cattle recognition system (shown in Fig. 1). The server
identification approaches include only invasive techniques. creates a database of cattle images. Figure 1 illustrates the
For example, embedded microchips or attachments (e.g., transferring of cattle images and recognition of cattle.
chips, RFID, transponders, ear-tagging, ear tip or notches, The muzzle point images of cattle are cropped from the
freeze-branding, and hot ironing) are invasive-based recog- extracted video frames [set of video frames (image)]. After
nition techniques. The benefits of the proposed cattle that image enhancement techniques are applied to mitigate
recognition system are cost-effective, non-invasive, auto- the noises and improve the contrast of muzzle point ima-
matic, easy to acquire, accurate and also humane. ges. However, the surveillance camera captures the video
In this paper, a muzzle point recognition-based solution from the unconstrained environment (e.g., low illumina-
gives the automatic recognition of cattle in the real time. tion, poor image quality, head movement of cattle, and
To address this problem, we use surveillance cameras for blurred). Moreover, muzzle point images are segmented to
the automatic capturing of images of cattle for identifica- find the discriminatory region of interest to extract the
tion purpose in the stables (farm house) of cattle. features from segmented muzzle point images.
For the recognition of cattle, surveillance cameras have After that muzzle features are obtained from the muzzle
been deployed in the stables of Dairy and husbandry, point image database, using proposed Fisher locality pre-
Institute of Agriculture Sciences (I.A.S) Bananas Hindu serving projections (FLPP) technique, texture feature
University (B.H.U), Varanasi-221005. The surveillance descriptor algorithms, and appearance-based feature
cameras continuously record and capture the images of extraction algorithms. The proposed system selects the sets
cattle. The proposed real-time cattle recognition system of discriminatory features from the extracted muzzle point
extracted the video frames (i.e., the set of pictures of cattle features. Then, unique templates of muzzle point image are
muzzle points pattern) from the captured surveillance video generated from the selected set of features and stored in the
and used as the input image for the cattle recognition muzzle point image database (shown in Fig. 1).
system.

Fig. 1 Illustration of transferring the images and recognition of cattle

123
 Author's personal copy
J Real-Time Image Proc

In the testing phase, a query (test) muzzle image of 5. The proposed Fisher locality preserving projection
cattle is matched with stored muzzle point template data- approach extracts the features of muzzle point images
base using one-shot similarity (OSS), distance matrices- of cattle. By applying this approach, only the most
based matching technique, and first-class SVM classifica- discriminant and stable muzzle features are preserved
tion model for the recognition of cattle in the real time for the representation and recognition of individual
(shown in Fig. 1). cattle.
The advantages of the proposed cattle recognition sys- 6. The database of muzzle point image pattern of 500 cattle
tem are low cost, invasive, automatic recognition, and (subjects) is prepared with surveillance cameras using
monitoring of animals in real-time scenario. The proposed 60–120 mm lens from the Department of Dairy and
real-time cattle recognition system does not need any Husbandry, Institute of Agriculture Sciences (I. A. S.),
expensive and noninvasive-based extra hardware, such as Bananas Hindu University (B.H.U.), Varanasi, India-
RFID-tags, embedded microchips, hot ironing, freeze- 221005. The size of muzzle image database is 5000 (i.e.,
branding, sensors, and the local database to recognize 500 subject’s 10 images of each subject). The size of
individual cattle. each muzzle point image is 200  200 pixels.
The proposed recognition system provides a non-contact
The rest of the paper is organized as follows: Sect. 2
solution, cost-effective, and robust recognition of livestock;
reviews the related works in the field of cattle recognition,
thus, there is no requirement to mount or embed any invasive
tracking, and monitoring in the real-time scenario. Sec-
artificial marking hardware identifiers to the cattle.
tion 3 illustrates the preparation and description of muzzle
point image database for recognition of cattle. Section 4
1.2 Major contributions of work
presents the proposed recognition system for cattle recog-
nition using their muzzle point images. Section 5 shows
The major contributions of this research work are as
brief descriptions of feature extraction and matching using
follows:
proposed Fisher locality preserving projections (FLPP) for
1. The unique physiological characteristics and un-coop- feature representation. Section 6 illustrates the experi-
erative behaviors of animal (especially cattle) achieve mental results along with their detail performance evalua-
new leads to interesting, significant challenges for tion and analysis. Finally, Sect. 7 is dedicated to the
cattle recognition using the animal biometric system in conclusion and future directions.
real time. Based on the consideration of non-intrusive
characteristics of animal biometric systems, this
research work explores the new possibility of cattle 2 Related work
recognition using muzzle point pattern recognition
algorithms for determining the identity of cattle based The animal biometric-based recognition systems have
on their primary biometric characteristics, such as received much attention in recent years for recognition,
muzzle point image pattern of individual cattle. tracking, and monitoring of different species or individual
2. In this paper, we propose the Fisher locality preserving animal. The recognition systems recognize the animal
projections approach for the recognition of cattle using using their physiological characteristics (e.g., facial ima-
muzzle point images in the real time and demonstrates ges, morphological pattern) and visual features of their
the current state-of-the-art results. body [9, 17, 20]. However, the prestigious animal recog-
3. The proposed real-time cattle recognition system nition-based systems recognize individual using identifi-
utilizes Fisher locality preservation technique for the cation methodologies [33]. The classical animal
recognition of individual cattle in the real time. The recognition methodologies can be divided into several
Fisher locality preserving projections are a new graph- groups, namely (1) permanent recognition methods, (2)
based muzzle point feature representation approach in semipermanent recognition methods, and (3) temporary
linear feature subspace that maximizes the inter-class recognition methods [35]. The classification of classical
(between-class) scatter of muzzle point image feature animal recognition approaches is shown in Fig. 2.
matrix and effectively minimize the intra-class The permanent animal recognition methodologies pro-
(within-class) scatter matrix of muzzle point images vide the animal protection by recognizing animal using ear
to improve the recognition accuracy. tattoos, ear-tip or notches, freeze-branding (hot iron around
4. We introduce a simple segmentation technique to find animals neck), and embedded microchips or transponders
the ROIs of muzzle point image pattern for the better in the electronic devices, sensors, and RFID [33].
segmentation of only required the most discriminatory For example, ear-tagging, ID-collar-based recognition
and informative area of muzzle point images. techniques are semipermanent-based recognition

123
 Author's personal copy
J Real-Time Image Proc

Fig. 2 Classification of classical animal recognition approaches

approaches for the identification of individual cattle. types of scenarios (1) active RFID-based animal identifi-
However, the embedded label of ear tags can be easily lost, cation and tracking technique and (2) passive RFID-based
duplicated, and forged. The ear-tagging systems include animal identification and tracking based technology. The
various kinds of metal clips and plastic tags. However, active RFID equipment consists of a temperature sensor,
these ear tags can cater different types of infections or the power source (battery backup) and microchips. It is also
critical diseases to animals after the embedded of these tags known as RFID transponder. The active tags can broadcast
into animal body the signal similar to a cell phone. Therefore, it sends the
For example, ear-tagging, ID-collar-based recognition alert message or signal to registered animal owners or farm
techniques are semipermanent-based recognition approaches house owner when livestock animal or cattle becomes sick.
for the recognition of individual cattle. However, the Moreover, it provides a way to easily track the health and
embedded label of ear tags can be easily lost, duplicated, and movement of the livestock animal in the herd. It also
forged. The ear-tagging systems include various kinds of provides a way to follow the animal in the real-time.
metal clips and plastic tags. However, these ear tags can On the other hand, passive RFID-based technique utilizes
cater various kinds of infections or critical diseases to ani- the various equipment and sensor devices. The owner or
mals after the embedded of these tags into animal body [3]. parentages of livestock animals mount these devices in the ear
The ear-tagging and ID-collar-based cattle recognition of different cattle. RFID readers are deployed in a lot of dif-
approaches are susceptible to damage, duplication, losses, ferent places for reading RFID unique number in the form
un-readable, and fraud of ear tags; therefore, these house or stables of animals [56]. However, RFID-based ani-
approaches do not perform well for identification of indi- mal identification framework requires a significant number of
vidual livestock cattle in the long term [59]. RFID readers, scanner, or various antennas for the proper
While in temporary recognition-based methodologies, it communications. Therefore RFID and sensor devices (GPS
recognizes the individual animal by applying the sketch systems, embedded wireless microchips) based animal iden-
pattern-based techniques. For example, paint or dying and tification and tracking approaches are not a cost-effective
RFID-based animal recognition with embedded transpon- solution for the cattle recognition and monitoring [46]
ders or sensors in their body are temporary identification All classical animal recognition approaches are invasive,
methods [29]. expensive and also vulnerable to losses of ear tags [31].
Among these available techniques, the RFID-based However, the performance of conventional animal recogni-
animal identification and tracking process is one of the tion methods is tiny, limited, due to their vulnerability to
most promising for the cattle or other livestock animals. losses, easily duplication, fraud, major security issues, and
The RFID-based animal recognition systems operate in two challenges [3]. Therefore, classical animal recognition

123
 Author's personal copy
J Real-Time Image Proc

approaches are unable to provide a competent level of The developed real-time-based framework system
monitoring and protection to livestock animals [51]. gathers the enormous amount of information by capturing
The research focus has shifted and advancement has the body variations in different poses from the top image
provided a new paradigm for the control, tracking, and sequence (video frames) of different surveillance video of
recognition of human or animals using surveillance camera pregnant cows in the real-time environment. The major
and intelligence devices [47] in the farm houses and live- shortcoming is that it does not validate the experimental
stock frameworks [16, 22, 65]. results during the classification of behavioral activities of
In the available literature, different models and frame- pregnant cattle before calving.
works are found to solve the problems of monitoring, In the direction of animal tracking, different tracking
tracking, recognition, and analysis of animal behavior. algorithms are developed [38]. These tracking approaches
Martiskainen et al. [44] proposed a system using computer offered more flexibility in the monitoring of large-sized
vision techniques for the monitoring and tracking of cattle animals in a constantly changing background. Author [37]
in the farm houses. The proposed system provides the proposed an animal recognition model using computer
tracking information of livestock and localizes the loco- vision system to study the behavior of hens in furnished
motion of postures of animals [11, 21, 52]. cages. Individual behavior, such as standing, walking, and
For example, various animal tracking and monitoring scratching, could be recognized automatically in real-time
techniques, such as global positioning system (GPS), Wi- scenario [10].
Fi, wireless networks, RFID, and Bluetooth-based moni- Cangar et al. [10] proposed a framework for the moni-
toring and tracking methods cater a better platform for toring and tracking of locomotion and posture, the behavior
tracking and monitoring of cattle [2, 25]. However, of pregnant cows, before calving, using computer vision-
deployments of these techniques are not suitable in practice based approaches in real time.
and more difficulties in attaching or embedding any elec- In the similar direction, health monitoring of poultry, the
tronic monitoring devices to cattle body. The classical author proposed a framework based system using computer
animal recognition and surveillance identification approa- vision approaches. They applied various surveillance
ches provide only evaluation and analysis, based on manual cameras for the diagnosis of critical diseases and moni-
scoring techniques for the quantification of animal behav- toring of poultry flock health [37]. Moreover, the authors
ior in the field or animal farmhouses [48]. proposed an animal recognition model using computer
The similarity matching-based techniques can apply for vision system to study the behavior of hens in furnished
the analysis of animal behavior using videos. These cages. Individual behavior, such as standing, walking, and
methods consume more time and require considerable scratching, could be recognized automatically in real-time
man-hour resources. Therefore, these techniques have scenario.
consumed more preprocessing and feature extraction time Shao et al. [55] proposed a method to identify and
for computing the feature vectors and predicting the correct analyze swine behaviors using the programmable cameras.
analysis of features. To provide the better improvement in The two parameters, such as area and the perimeter of the
the behavior analysis, expert as human interpretation is also top views, were applied to detect the swine from the cap-
required as individual support [41]. tured image database.
To solve these major problems various intelligence In the same direction, Tillett et al. [57] used the image
devices, sensors, and chip-based devices have been processing and computer vision-based techniques for the
deployed by the embedding of these sensors and devices to recognition and analysis of the behavior of pigs. The pigs
the animals body for the quantification and analysis of were identified in surveillance video by tracking and
animal behaviors. However, these methods are also inva- detecting their movements.
sive for animals and alter the behavior of the animal [26]. Recently, in commercial livestock houses, image analysis
The computer vision, machine learning, and image for behavior classification becomes more complicated due to
processing techniques have achieved recently more atten-
tion for recognition and monitoring of cattle in the dairy
Table 1 Details of the muzzle point image pattern database
industry [12, 42].
The author [28] proposed a framework for the investi- Breeds (races) No. of subjects (cattle) No. of images
gation of locomotion of dairy cows using high-speed cin- Balinese cow 150 1500
ematography techniques [18]. The identification Hybrid Ongole cow 150 1500
framework and computational models are employed in the Holstein Friesian cow 100 1000
variety of applications for cattle recognition such as Crossbreed cow 100 1000
tracking of livestock [43].

123
 Author's personal copy
J Real-Time Image Proc

better research and experimentation purpose. However, few

researchers have proposed some cattle recognition
approaches to identify individual cattle based on only their
muzzle (nose) print image databases of cattle.
The beads pattern of muzzle point image consists of
irregular structures, and their shape is similar to the islands,
whereas ridges pattern is similar to rivers structure. The
ridge pattern always separates the beads. The beads and
Fig. 3 Some muzzle point image pattern of cattle from database
ridges of muzzle point are unique biometric identifiers for
the recognition of individual cattle. Therefore, these animal
lighting condition and dynamic changes. But camera char- recognition systems are not available in the public domain
acteristics, background, and test subjects characteristics all for the recognition and verification of cattle.
these things influence the ability of the system to recognize
the issue and record its movement accurately [24].
4 Proposed cattle recognition system

3 Database preparation and description In this section, the proposed real-time cattle recognition
system using muzzle point images is illustrated in detail.
The database of muzzle point image is prepared with For the manifestation of the objective of the proposed cattle
surveillance cameras using 60–120 mm lens from the recognition system, the surveillance cameras are deployed
Department of Dairy and Husbandry, Institute of Agriculture in the cattle stables of Department of Dairy and Husbandry
Sciences (I. A. S.), Bananas Hindu University (B.H.U.), for continuously monitoring through capturing the animals
Varanasi, India-221005. The size of muzzle point image (especially cattle) video.
database is 5000 (500 subject 10 images of each subject). The proposed real-time cattle recognition system con-
The size of each muzzle point image is 200  200 pixels. sists of two phases: (1) training phase and (2) testing phase.
The detail description of the database is given in Table 1 In the training phase, the cattle recognition system builds a
and Fig. 3 illustrates some muzzle point image pattern of database of the muzzle point images from the surveillance.
cattle. The proposed system extracts the video frames (i.e., a set
of images) from the captured surveillance video.
3.1 Characteristics of muzzle point image pattern The obtained cattle images are applied as input images
of cattle to cattle recognition system. The muzzle point images are
cropped from the extracted video frames. The image
In our affirmation, real-time cattle recognition system using enhancement techniques are applied to mitigate the noises
muzzle point image pattern can be more friendly, noninva- and improve the contrast of muzzle point images because
sive, and cost-effective for identification, verification, moni- the video is captured from the unconstrained environment
toring, and tracking of individual animals, if the performance (e.g., low illumination, poor image quality, head movement
of automatic matching algorithms is more satisfactory. of cattle, and blurred).
The muzzle point image is a unique and stable discrim- The muzzle point images are preprocessed to mitigate
inatory biometric characteristic of cattle for the recognition the noises, and images are converted into grayscale muzzle
purpose. The recognition of muzzle point image pattern is images. After that, muzzle point images are segmented to
similar to the recognition of minutiae points in the human find the discriminatory region of interest to extract the
fingerprint [3, 27, 34]. features from segmented images.
The discriminatory biometric pattern of muzzle point After that, muzzle features (e.g., texture features and
images can be clustered into two significant patterns known pixel intensity values of muzzle point images) are extracted
as beads and ridge pattern. The beads pattern of muzzle from the muzzle point image database using texture feature
point image consists of irregular structures, and their shape descriptor algorithms, appearance-based feature extraction
is similar to the islands, whereas ridges pattern is similar to algorithms. The discriminatory muzzle features are selec-
rivers structure. The ridge pattern always separates the ted from the extracted features. The unique templates of
beads. The beads and ridges of muzzle point are unique muzzle point image are generated from the selected set of
biometric identifiers for the recognition of individual cattle. features. Then, muzzle image templates are generated and
The research in the field of cattle recognition using their stored in the muzzle point database.
muzzle point images characteristics is limited by the non- In the testing phase, recognition system takes the input
availability of a large muzzle point image database for the muzzle image as a test (query) image. The muzzle point

123
 Author's personal copy
J Real-Time Image Proc

Fig. 4 Proposed block diagram of real-time cattle recognition system

features are extracted using proposed graph-based feature

extraction technique known as Fisher locality preserving
projections (FLPP).
The FLLP feature extraction and representation tech-
nique utilizes two feature graphs with the first graph uti-
lized to distinguish the compactness of within-class ðSW Þ
scatter matrix of muzzle point image database. The second
graph is assigned and employed to the augmentation of
scatter matrix feature separability of the between-class ðSB Þ
in the muzzle point image database.
The primary evidence to apply the proposed FLPP
approach is to preserve only the most discriminant and
immutable muzzle point image features retained. Finally, the
feature vectors of muzzle point images are achieved and
matched with their counterpart in the stored muzzle point Fig. 5 Illustration of pattern of muzzle point image for the calf
templates databases. After that, obtained similarity scores of registration

123
 Author's personal copy
J Real-Time Image Proc

Fig. 6 Beads and ridges of the

muzzle point pattern of cattle
from database

unconstrained environments (i.e., poor illumination, pose,

and blurring of pictures) using surveillance video. There-
fore, it may be defective and deficient in some respect, such
as poor image quality, low contrast, and blurred. Some
sample images of muzzle point are shown in Fig. 7. The
preprocessing steps are essential for feature extraction,
segmentation, feature representations, and matching.
The segmentation technique is applied to find the ROI
from the original muzzle point images and provides
enhancement of the recognition accuracy of cattle. How-
ever, the ROI of muzzle images must be chosen carefully
to cater all the discriminatory features of muzzle point
images. The preprocessing step is shown in Fig. 8.
We have applied a CLAHE image enhancement technique
to improve the contrast of low quality of muzzle point images
[50]. Enhancing the muzzle image contrast helps to reduce the
Fig. 7 Some muzzle point image of cattle has problem of poor
illumination effect and boost the fine patterns of beads and
illumination (1, 3, and 6) and blurred muzzle point images (2, 4, and
5) during database acquisition rides as main discriminatory features are shown in Fig. 8.
Figure 8 illustrates the pre-processing of beads and the ridge
muzzle point images are compared with defined threshold pattern of original muzzle point images. In Fig. 8, (a) shows
values ðhÞ to produce a final decision for cattle recognition. the input muzzle point image, (b) illustrates the blurred
The proposed real-time cattle recognition system is muzzle point image pattern, (c) demonstrates the pre-pro-
shown in Fig. 4. The steps involved in the real-time cattle cessing and extraction steps for the beads and the ridge pattern
recognition are given in brief as follows: of muzzle point images, and (d) shows the filtrating process of
Figures 5 and 6 illustrate beads and the ridge pattern of overlapped regions between beads and ridge pattern of
muzzle point images. The bead and ridges are extracted for the muzzle point images. After the enhancement of the low
recognition of cattle using texture descriptor-based recogni- illumination, poor image quality, and blurred muzzle images,
tion techniques and appearance feature extraction algorithms. we have applied the segmentation techniques to partition the
muzzle point images into different image segments to find out
4.1 Preprocessing of muzzle point images the region of interest. After that, segments of muzzle point
images are selected to locate the region of interest (ROI) in the
In the preprocessing steps, preprocessing techniques are muzzle point images. Each region of interests of muzzle point
applied to mitigate the specific degradation, such as noises images is determined for the extraction of discriminatory
from captured muzzle point images. The muzzle point features (the most informative pattern of beads and ridge area)
images are from captured video which are taken from of muzzle point images.

123
 Author's personal copy
J Real-Time Image Proc

Fig. 8 Preprocessing of beads

and ridges from the original
images of muzzle point pattern

Fig. 9 a Input muzzle point

images b segmented of muzzle
point pattern using a K-means
clustering segmentation
algorithm

4.2 Segmentation of muzzle point images segmented muzzle point images using color K-means
clustering and texture-based segmentation technique are
In this section, a simple muzzle image-based ROIs seg- shown in Figs. 9, 10, and 11, respectively.
mentation algorithm for muzzle point image is illustrated in Figure 11 illustrates the segmentation of muzzle point
brief. The segmentation algorithm performs the segmen- image, where ground truth images are shown in (b–e) to
tation of muzzle point images into the number of ROIs find the ROIs from the segmented original muzzle
muzzle bead and the ridge pattern for the extraction of only image, where (a) shows the segmentation of muzzle
the most discriminatory and more informative feature from images using texture segmentation algorithm. The muz-
these segmented RIO areas. zle point images contain discriminatory color features.
After the enhancement of muzzle images, segmentation Therefore, a K-means color-based segmentation tech-
algorithms are applied to find ROIs from the segmented nique is applied to segment the texture feature into
muzzle point images. We have applied color K-means different clusters, where (a) illustrates the segmentation
clustering and texture-based segmentation algorithms. The of muzzle images using texture segmentation algorithm

123
 Author's personal copy
J Real-Time Image Proc

Fig. 10 Segmentation of
muzzle point image using
K-means color segmentation
technique

Fig. 11 Illustrates segmentation of muzzle point image

and (b) segmented muzzle point pattern, shown in reduced image quality, low contrast, and blurred. There-
Fig. 10. fore, existing feature extraction, representation, and simi-
larity matching algorithms are unable to perform the
4.3 Learning-based encoding and distance metric recognition of cattle using muzzle point images [7].
approach To achieve the recognition of cattle, learning-based
feature extraction and matching algorithms are essential to
As depicted in the previous section, the captured muzzle cater an explicit encoding mechanism of feature space to
point images have several significant challenges due to the improve the recognition accuracy of individual cattle in the
unconstrained environment, such as poor illumination, real-time scenario. Therefore, the proposed cattle

123
 Author's personal copy
J Real-Time Image Proc

recognition approach is motivated by feature representa- pairs of elements or instances from a set S. It can also apply
tion, learning-based feature extraction, and distance metric a distance measure function for learning and classifying
learning-based matching of muzzle point feature vectors. given instances [54, 64].
In this subsection, learning based on distance metrics One-shot similarity (OSS) matching is a semi-super-
one-shot similarity (OSS) using Fisher linear discriminant vised-based matching similarity technique. It selects the
analysis (FLDA), and learning-based distance matrices via unlabeled training data as a set of negative constraints
OSS similarity matching technique with first-class online against which two input sample images of muzzle point
incremental SVM classification models are applied in the pattern are matched [61].
proposed real-time cattle recognition system. The OSS-based distance function measures the dissim-
The 1-class online incremental SVM (1-online ISVM) ilarity between the given case of muzzle point images and
model is used to classify the extracted texture features of separates the class of negative instances of muzzle images
muzzle point images. We have done the customization in the that are presented during the matching process of muzzle
batch formulation of 1-class online incremental SVM (1-on- point images in the testing phase of proposed cattle
line ISVM) model. The formulation customization provides a recognition system. The OSS-based matching of distance
method to increase or decrease the sample images of muzzle between the classes is computed using the classifier shown
point pattern for the better learning the model and classify the in Fig. 12. The output is obtained using each muzzle
extracted texture features of muzzle point images. The cus- instance and negative background set (L).
tomization caters the number of available incremental sam- Generally, during the matching of a given pair of muzzle
ples of muzzle images for recognition of cattle. image’s instances, let (A) and (B) are the two instances of
The major shortcoming of second-class-based SVM clas- the muzzle point images. First, an instance-specific model
sification and learning model is that it always requires many of SVM classifier is trained with instance (A) as a single
labeled sample of muzzle point images from the two classes. positive sample image of muzzle point and a separate set of
while first-class SVM classification and learning model muzzle point image (L). The sample images are not (A) and
is faster and performs better classification of extracting (B) as shown in Fig. 12 of negative samples. A similarity
features than second-class SVM classification and learning matching-based index between (A) and set (L) is computed
model. However, it is more suitable for incremental update as ðS1 Þ. Similarly, a similarity matching index is trained
of the database in real time [15]. and commuted between (B) and the same set (L) as ðS2 Þ.
The OSS technique computes the similarity between the
4.3.1 One-shot similarity (OSS) using FLDA set of scores ðS1 Þ and ðS2 Þ using the summation rule.
Therefore, the formulation based on Fisher linear dis-
A metric function was defined as a definite positive dis- criminant analysis (FLDA) is applied as the supervised
tance measurement function for the calculation of distances classifier technique to classify and recognize the extracted
between two elements of a given set. For a given set S, features. The availability of only one element in the posi-
metric function ðDÞ : S  S ! ½0; 1 is defined. It carries tive class restricts the number of possible classifiers that
out distance function that applied to measure the dissimi- can be used [36].
larity between two elements in the given sets S [54]. The FLDA classification technique is used for learning
On the other hand, classical distance-based metric distance matrix to obtain the advantage of pre-computation
measure functions, such as Euclidean distance function and of covariance matrix from all muzzle point image pattern
Chi-square distance measure ðv2 Þ learning-based metric of within-classes [62]. The computation using FLDA
expostulate only a set of discrete binary (0 or 1) labeled classification technique is given in brief as follows:
Fig. 12 One-shot similarly
(OSS)-based matching for A and
B two instances of muzzle
images is matched

123
 Author's personal copy
J Real-Time Image Proc

Let Si 2 Rd where i ¼ ð1; 2; 3;    ; N Þ is the set of pos- The proposed cattle recognition approach is motivated
itive sample images of muzzle point and Mj 2 Rd , where from the feature extraction and representation-based learn-
j ¼ ð1; 2; 3;    ; KÞ is the set of other classes of sample ing and distance metric learning-based matching techniques.
images of muzzle point with mean values lr ls and l For this research-based experimentation, the database of
(mean of all points in the muzzle images classes), respec- muzzle point of cattle contains 5000 muzzle images from
tively. According to [64], between-class ðSB Þ and within- 500 cattle. Each cattle has 10 images with 200  200 pixels.
class ðSW Þ of the muzzle point image database are defined For the experimentation and encoding purpose, the
as follows (shown in Eqs. 1 and 2): database of muzzle images is grouped into two major
classes, such as between-class ðSB Þ and within-class ðSW Þ
SB ¼ ðlr  ls Þðlr  ls ÞT ð1Þ of muzzle point image database. For the better recognition
XN rate for identifying individual cattle, there is a need to
1
SW ¼ ðPi  ls ÞðPi  ls ÞT þQ ð2Þ provide a better separability of extracting features of
ðN þ K Þ i¼1
muzzle point images among these classes.
1
P T To achieve better separability between classes, we pro-
where Q ¼ ðNþKÞ j¼1 ðKðMi  ls Þ  ðMi  ls ÞÞ :
pose Fisher locality preserving projections (FLPP)
The FLDA technique first computes a linear projection approach for the maximizing the between-class scatter by
u. It maximizes the scatter matrix separation of between minimizing the within-class scatter of the muzzle point
classes of muzzle images by minimizing the within classes. image database. The proposed FLPP approach is illustrated
The computation is given as follows: in detail in the next subsection.
uT SB u
WOPT ¼ argmaxu ð3Þ
uT SW u
5.1 Proposed FLPP feature extraction
In the linear projection, L keeps a single sample of muzzle and representation approach
images (e.g., (A) or (B)) and N is a separate set classes of
muzzle image. For the recognition of cattle using muzzle The proposed FLPP approach determines the directions
image (x), a similarity score is measured as uT ð xÞ  u0 , where that maximize the between-class scatter matrix ðSB Þ and
ðu0 Þ is defined as bias and given as u0 ¼ ðuT Þ  ðxþl rÞ
2 .
maintain the local geometrical data in feature space. These
Therefore, similarity between scores is measured and com- advantages provided us motivation to propose an improved
puted from these set of sample image (background (L) sam- FLDA-based feature extraction and representation of
ple) using OSS learning and distance metric technique can muzzle point images. The extracted features of the muzzle
improve the matching performance for cattle recognition in point image are represented using proposed feature repre-
real-time scenario [(L) in the example shown in Fig. 12]. sentation approach known as Fisher locality preserving
projections (FLPP).
The proposed FLPP approach utilizes two feature
5 Feature extraction and matching image-based graphs for the representation of the muzzle
point feature. The first feature graph is used to characterize
The features are set of pixel intensity values of the given the within-class ðSW Þ scatter matrix of muzzle point images
object in the image. The large set of pixel values is encoded compactness, and the second graph is dedicated to the
and transformed into a reduced set of pixel values for augmentation of the between-class ðSB Þ separability.
representation of images. The reduced set of pixel values is The proposed FLPP approach determines the global
known as feature vectors. distribution of within-class ðSW Þ scatter and between-class
As motioned in the previous section, cattle recognition ðSB Þ scatter matrix of the muzzle point image database
using muzzle point images has significant challenges that locally, such that class separability and neighborhood
arise from their continuous movement of the head, pose, structure preservation of muzzle image database are
posture, and deformation of animals body during data obtained simultaneously.
acquisition. Therefore, existing feature handcraft extrac- The primary motivation of the proposed FLPP approach
tion, representation, and matching techniques are unable to is to maximize the between-class ðSB Þ scatter by mini-
perform the better recognition in feature space [8]. mizing the within-class ðSW Þ scatter of the muzzle point
In such scenario of the unconstrained environment, image database. The objective of FLDA is illustrated as
learning-based feature extraction, representation, and follows (shown in Eqs. 4–9):
matching algorithms are essential to perform the encoding X
N
of the extracted feature of muzzle point images and to SB ¼ ðNi ðmi  mÞ  ðmi  mÞÞT ð4Þ
i¼1
provide a better recognition rate of cattle in the real time.

123
 Author's personal copy
J Real-Time Image Proc

N X
X matching-based learning, distance matrices-based match-
SW ¼ ðNi ðmi  mÞ  ðmi  mÞÞT ð5Þ ing, and incremental first-class incremental SVM model for
i¼1 xk
the classification and recognition of muzzle point images of
where SB , SW and l are defined as a between class, within cattle [15]. The proposed FLPP approach is applied for
class, and mean of classes of muzzle image database, maximizing the class separability using graph-based rep-
respectively. resentation, shown in Algorithm 1.
1X N
l¼ Ni ð6Þ
N i¼1

uT SB u
WOPT ¼ argmaxw ð7Þ
uT SW u
In many cases, within class is singular; therefore, there is in
need to solve the singular equivalent construction can be
applied:
u T SB u
WOPT ¼ argmaxw T ð8Þ
u ðSW þ SB Þu
uT SB u
WOPT ¼ argmaxw ð9Þ
uT ðSR Þu
where SR is defined as total scatter matrix and can be
computed as follows:
c X
X   T
SR ¼ xj  li  xj  li ð10Þ
i¼1 xj 2Xi

In the above Eqs. (4–9), the total scatter matrices depend on

the total mean value of extracting features not on the class
means. Therefore, it is computed without having any
information of class labels. The total scatter matrices can be
illustrated in a generalized Eigenvalue problem as follows:
SB V ¼ KSR V ð11Þ
where Eq. (11) generalizes the solution V of the Eigenvalue-
based problem. It keeps the most discriminant projection
direction in the feature space, where c and n are the total
number of sample images and the number of classes of
muzzle point image pattern, ðni Þ is the number of samples in
the ith class. Moreover, ½ X  ¼ ½X1 ; X2 ;    ; Xn  is the data
(feature) matrix. The data matrix contains the extracted bead
and ridge features of muzzle images, ðXi Þ is the set of samples
related to ith class of muzzle images. ðli Þ is the mean of
training samples of the muzzle point image database.
To achieve the maximum separability between the
classes of muzzle point images, initially, the two graphs are
constructed and features are represented using FLPP
algorithm. The representations of muzzle point feature
based on graphs are shown in Algorithm 1. At the begin-
ning, the set of n muzzle point images are defined as X ¼
½X1 ; X2 ;    ; Xn  in RðndÞ .
In this paper, we assumed that each muzzle point image
belongs to one of the c classes, such as x ¼ ½x1 ; x2 ;    ; xn . The proposed FLPP outperforms the conventional Fisher
The proposed FLPP algorithm applies the OSS similarity linear discriminant analysis (FLDA) and locality

123
 Author's personal copy
J Real-Time Image Proc

Fig. 14 Representation of scatter plot of 10 different pixel value of

Fig. 13 Scatter plot of 10 different pixel value of muzzle classes with muzzle point image classes with 250 samples image of muzzle images
250 samples image of muzzle images using the LLP and FLDA using proposed FLPP approach
technique

preserving projections (LPP) methods. For example, in the algorithm are applied for the compartive study of experi-
case of LPP, FLPP provides a superior compaction repre- mental results. The brief description about the texture
sentation of the within-class ðSW Þ muzzle images by min- feature extraction and appearance-based feature extraction
imizing the distances between-class ðSB Þ data (features) and representation techniques are discussed in brief in the
and its neighbors of the same class. The representation of next subsection.
10 different data classes in the scatter plot of muzzle
images is shown in Figs. 13 and 14, respectively. 5.2 Appearance and texture-based feature
extraction algorithm

Appearance-based face recognition algorithms are classical

feature extraction and data representation technique. It is
widely used in the area of pattern recognition and computer
vision.
In this paper, we have applied the appearance-based
feature extraction and representation approaches for the
recognition of individual cattle. The used appearance-based
approaches are, namely, principal component analysis
(PCA), PCA-LiBSVM [6, 58], linear discriminant analysis
(LDA), independent component analysis (ICA) [4, 39]
ICA-LiBSVM, Batch-ILDA [30] and its modified version
algorithms (i.e., Batch-CCIPCA [60]), LDA-LiBSVM,
ILDA-LiBSVM [30] and local projection preserving
(LPP)-based approaches for recognition of cattle
[19, 23, 53].
The feature of muzzle point images consist of dense skin
texture information. The texture features are varying illu-
In this paper, for the comparative analysis of experi- mination and poor image quality under unconstrained
mental results, texture feature-based descriptor techniques environment. For the extraction of texture features from
and appearance-based face recognition method are used. muzzle images, speeded up robust features (SURF) [5],
The feature descriptor methods namely speeded up robust local binary pattern (LBP) [1, 49], and scale invariant
feature (SURF), local binary pattern (LBP), and appear- feature transform (SIFT) [40] algorithms are applied.
ance-based recognition approaches (e.g., principal com- Therefore, extracted texture features are encoded into a
ponent analysis (PCA), linear discriminant analysis (LDA), binary pattern for the better matching of muzzle images
independent component analysis (ICA) and their various and enhancement of recognition accuracy.

123
 Author's personal copy
J Real-Time Image Proc

5.3 Matching of muzzle point images suppress of extreme points of muzzle point from the ROIs
of muzzle point images are shown in Fig. 16.
In this subsection, matching of muzzle point images using
learning distance-based matrices matching technique, OSS
via SVM, and learning distance matrices via OSS with 6 Experimental result and analysis
first-class incremental SVM classification techniques is
given in detail [61]. The advantage of incremental SVM In this section, the experimental results are performed on
classification model is to provide updates regularly to adapt inlet core 2 duo GHz computer with 40 GB of RAM. First,
the variations in pixel intensity in the extracted features of database of muzzle point image is cropped and re-sized
muzzle images [15]. into 200  200 pixels from the frontal face images of each
Further, a customization in the batch-based learning cattle. After preprocessing, enhancement, and segmenta-
formulation of first-class SVM model and Incremental tion of muzzle point images, pixel intensity and texture-
SVM (I-SVM) model for the incremental or decremented based muzzle point image features are extracted from the
(online) learning procedure is done. It caters a sufficient database using texture feature-based descriptor techniques.
availability of an incremental database of muzzle point
image with time. These techniques successively eliminate a 6.1 Performance evaluation
number of histories of samples and restore several new
samples. The matching of two muzzle images (based on For performance evaluation of experimental results, the
minutiae feature points) is shown in Algorithm 2. first muzzle database was segmented into following parts:
Figure 15 illustrates the detection of minutiae points and (1) gallery (training) part and (2) testing part. For the
inverted skeleton image with core point (green color), experimental results, six muzzle images of each cattle
lower core points (gold color), bifurcations (blue color for (subject) (i.e., 500 cattle 6 muzzle images of each cattle =
h 2 ½0  180  and purple for h 2 ½180  360  and ridge 3000) muzzle images are applied for the training purpose
endings (orange color for h 2 ½0  180  and red color for of proposed real-time cattle recognition system. For the
h 2 ½180  360 . The finding of the segmented ROIs and testing phase, four muzzle point images are used as query
image for the matching with stored trained muzzle image
template for the recognition of individual cattle.
The evaluation of rank-1 recognition accuracies is
computed and validation of accuracy is completed on five
times crossvalidation in this experiment. For the compu-
tation of recognition accuracy, texture feature-based
descriptors algorithms, such as SURF [5], LBP [1, 49] and
SIFT [40] and appearance-based face recognition algo-
rithms (PCA [58], LDA [19], ICA [4, 39] and its variants)
are applied to the feature extraction and representation of
muzzle images in feature space.
Fig. 15 Detection of minutiae points of muzzle images The experimental results are illustrated in the form of
cumulative match characteristics (CMC) curves. The CMC

Fig. 16 Illustration of segmentation process of muzzle point images to find the ROI

123
 Author's personal copy
J Real-Time Image Proc

curve illustrates the identification accuracy of recognition number of choosing LDA Fisher component feature vectors
systems, how well this system discriminates individual by varies from 10 to 50. The number of selected PCA com-
identifying in the enrolled muzzle point image database ponent varied from 50 to 200 in intervals of 10. Based on
with respect to a given test or probe muzzle image of cattle. the chosen features of muzzle point image pattern, the
The recognition accuracy of cattle presented in form of nearest neighbor classifier is applied to classify the muzzle
CMC curve, which is shown in Figs. 17, 18 and 19, features. The recognition accuracy of cattle is illustrated in
respectively and accuracy is illustrated in Tables 2, 3 Fig. 17.
and 4, respectively. In this experiment, the performance of The LDA algorithm performs better than PCA tech-
the proposed approach is evaluated, and its comparison is nique. The top recognition accuracy of LDA algorithm is
performed with other existing face recognition algorithms. 79:95 and minimum recognition accuracy 75:57 using 50
The set of salient features is extracted using the feature vectors; however, PCA techniques yield 75:86
appearance-based face recognition algorithms, such as using Eigenfaces with 200 feature sets, because PCA-based
PCA, LDA, ICA, and its modified variant algorithms, Eigenface technique fails to perform the muzzle point
respectively. Then, 200 PCA component features and 200 recognition in low illumination and poor image case. It
Fisher discriminant feature vectors (sets) were obtained achieves the minimum number of principal components
using PCA and LDA feature extraction techniques. The (i.e., maximum variance feature) of extracting muzzle

Fig. 17 Recognition accuracy of cattle using PCA, LDA, ICA, LBP, Fig. 19 Recognition accuracy of cattle using Batch-iLDA, CCIPCA-
SURF, and proposed approach LiBSVM, ICA-LiBSVM, iLDA, and iLDA-LiBSVM algorithm

Fig. 18 Recognition accuracy

of cattle using Batch-CCIPCA,
ICA, IND-CCIPCA, ISVM,
LDA, LDA-LiBSVM, PCA, and
PCA-LiBSVM

123
 Author's personal copy
J Real-Time Image Proc

Table 2 Recognition accuracy (%) of cattle using PCA, LDA, ICA, algorithms such as PCA, LDA, ICA, SURF, LBP, and
SURF, LBP, and proposed approach proposed FLPP algorithms for the recognition of muzzle
F-sets PCA LDA ICA LBP SURF Prop. point image pattern of cattle. The recognition accuracy is
amplified by increasing the number of feature sets of
50 74.39 75.57 86.97 78.68 83.40 86.67
muzzle images.
100 79.81 80.64 79.92 82.20 62.10 89.78
150 81.89 84.19 78.97 85.92 60.95 93.83 Table 3 illustrates the recognition rates of Batch-
200 75.86 79.95 87.95 92.95 94.57 96.87 CCIPCA, ICA, IND-CCIPCA, ISVM, LDA, LDA-
Prop. proposed approach, F-sets feature sets LiBSVM, PCA, and PCA-LiBSVM for the recognition of
cattle. The matching using one-shot similarity (OSS)-based
point image features. It yields minimum recognition rates learning technique and learning distance-based matrices
as compared to the LDA and ICA recognition approaches. approach is incorporated with SVM classification model.
As shown in Table 2, independent component analysis The customized batch formulation in the Incremental SVM
(ICA) feature representation technique exploits only (ISVM) is also applied with PCA and LDA for incremen-
higher-order dependencies in the pixel intensity values of tally updating the sample images of muzzle images.
muzzle point images. ICA method mitigates the first and Therefore, the recognition accuracy is achieved by ISVM
second-order statistical data (pixel values) by sphering the technique which is higher than other algorithms because
data on the computation of covariance matrix. In this the previous history of sample images is regularly updating
experiment, appearance-based feature extraction and rep- the incrementally added sample images online.
resentation approaches are also tested on muzzle point The recognition rates of muzzle point images using
images captured in the unconstrained environment (e.g., Batch-iLDA CCIPCA-LiBSVM,ICA-LiBSVM, iLDA, and
poor illumination and blurriness). Based on observations iLDA-LiBSVM algorithm are shown in Table 4. Fig. 19
ICA technique performs better recognition of cattle in the shows the identification accuracy of batch-CCIPCA,
unconstrained environment. It yields the higher recognition incremental support vector machine (ISVM), LDA-
accuracy as compared to PCA (eigenfaces) and LDA face LiBSVM, PCA, and PCA-LiBSVM for recognition of the
recognition and representation techniques because ICA muzzle point pattern of cattle. It shows that the ISVM
method computes the muzzle point features and selects algorithm yields 86.98 % recognition accuracy on other
only independent features from the extraction features by algorithms.
illuminating the first and second-order statistical data (pixel In Batch-CCIPCA technique, the labeled training sam-
intensity values) of muzzle point images. ple images of the muzzle are not available beforehand, it
The ICA technique performs the computation on pixel seeks the required training data and achieved it incre-
intensity to obtain the higher variance of extracting fea- mentally. However, in incremental steps, it selects a min-
tures. Figure 17 illustrates the performance of recognition imum number of features for updating the previous data.

Table 3 Illustrates the performance of modified appearance algorithms such as Batch-CCIPCA, ICA, IND-CCIPCA, ISVM, LDA, LDA-
LiBSVM, PCA, and PCA-LiBSVM
F-sets Batch-CCIPCA ICA IND-CCIPCA ISVM LDA LDA-LiBSVM PCA PCA-LiBSVM

Recognition accuracy (%)

50 66.67 82.75 50.95 82.40 73.33 74.29 60.25 64.78
100 70.49 84.29 54.32 87.68 76.79 79.95 63.75 68.82
150 74.95 86.34 58.95 90.98 80.95 87.59 66.85 71.86
200 75.86 88.95 62.75 94.95 84.57 91.57 68.58 74.98

Table 4 Recognition rates (%)

F-sets Batch-iLDA CCIPCA-LiBSVM ICA-LiBSVM iLDA iLDA-LiBSVM
of muzzle point images of cattle
based on Batch-iLDA, Recognition accuracy (%)
CCIPCA-LiBSVM, ICA-
LiBSVM, iLDA, and iLDA- 50 73.49 74.95 75.79 80.75 70.93
LiBSVM 100 77.78 78.93 82.95 84.97 74.92
150 82.85 86.95 88.95 90.95 78.25
200 86.98 89.95 92.95 94.95 85.57

123
 Author's personal copy
J Real-Time Image Proc

Therefore, the previous database is updated with minimum Cattle Image 1

features. After that, it selects only existing features. Cattle Image 2
Therefore, Batch-CCIPCA technique yields minimum Cattle Image 3
Existing work on Image 1

Success rate in %.
recognition accuracy. Existing work on Image 2
The recognition accuracy of independent-CCIPCA 100 Existing work on Image 3
increases with increasing set of features. However, the
80
number of selected eigenvalues of muzzle point image
decreases in each level of feature sets. Therefore, the 60
recognition accuracy of the IND-CCIPCA technique is low 40
as compared to other algorithms. The IND-CCIPCA algo-
20
rithm leaves top 10 muzzle point image pattern which
shows a minimum variance of extracted sets of features 0
0 20 40 60 80 100
(e.g., pixel intensity of the muzzle image pattern). The Data size (image size) in kB.
iLDA-LiBSVM recognition technique provides 85:57
recognition accuracy. The iLDA technique yields 94:95 Fig. 20 Impact of recognition time for different sizes of cattle images
recognition accuracy which is higher than other used for recognition of cattle
techniques due to the selection of an updates maximum
number of features at each level of feature sets. Figure 20 illustrates the impact of recognition time for
different muzzle image of cattle. The variation in the data
6.2 Impact of recognition time for different cattle size of cattle images at the required time of recognition is
images increased in different sizes of cattle images.
In Fig. 20, as the size of image increases, the cattle
In recognition of cattle, we demonstrate the significant recognition system needs more time to process and
impact of cattle images on the required time for recognition matching of muzzle features for recognizing and verifying
of cattle (e.g., success rate) of the cattle recognition system. the individual cattle.
The cattle recognition time consists of time taken for In the process of identification, the system performs
processing of cattle images and matching time during several preprocessing and feature extraction and similarity
training and testing of cattle images for the recognition and matching of muzzle point features. Therefore, in the cattle
verification of cattle. recognition system, the large size of data also takes more
In this result, surveillance camera performs first by time for image preprocessing, enhancement, and feature
detecting the cattle and capturing the image of cattle. After extraction. It observed the success rate (recognition rate) of
that, captured images (different size of cattle images) are the proposed cattle recognition system using cattle image-3
wirelessly transferred to the server of the cattle recognition [13].
system. The recognition time of muzzle images for iden-
tifying cattle in the real-time scenario is shown in Table 5.
7 Conclusion and future directions

Table 5 Recognition time for different cattle images using muzzle In this paper, we proposed a muzzle point-based cattle
point recognition approach for cattle recognition recognition system for identifying the individual cattle in
Image Image size (kB) Recog. time the real-time scenario. The proposed recognition system
captures the images of cattle by the surveillance camera.
Recognition time (s) The system extracts the video frames from the captured
Cattle image-1 10 11.50 video of the surveillance camera. The images of cattle are
Cattle image-2 20 13.25 transferred to the server of the cattle recognition system by
Cattle image-3 40 14.85 wireless network (internetwork) technology. This research
Cattle image-4 50 20.98 presents a current state-of-the-art approach for the recog-
Cattle image-5 60 21.78 nition, verification, and monitoring of cattle using
Cattle image-6 70 22.98 surveillance camera based on their muzzle point image
Cattle image-7 75 25.68 pattern in the real time.
Cattle image-8 80 26.98 A comprehensive description of proposed Fisher linear
Cattle image-9 95 27.99 preserving projection (FLPP) approach is applied for
Cattle image-10 100 30.74 extraction and representation of features of muzzle point
Recog. time recognition time (s) images. The extracted muzzle features are matched using

123
 Author's personal copy
J Real-Time Image Proc

one-shot similarity and incremental first-class SVM learn- 5. Bay, H., Ess, A., Tuytelaars, T., Van Gool, L.: Speeded-up robust
ing model for the recognition of cattle. The proposed features (surf). Comput. Vis. Image Underst. 110(3), 346–359
(2008)
approach yields 96:87 recognition accuracy. 6. Belhumeur, P.N., Hespanha, J.P., Kriegman, D.J.: Eigenfaces
With the appearance (holistic)-based face recognition verdsus fisherfaces: recognition using class specific linear pro-
and representation algorithms, independent component jection. IEEE Trans. Pattern. Anal. Mach. Intell. 19(7), 711–720
analysis (ICA) recognition technique yields 87:95 of (1997)
7. Bharadwaj, S., Bhatt, H.S., Vatsa, M., Singh, R.: Domain specific
recognition accuracy. On the other hand, texture descrip- learning for newborn face recognition. IEEE Trans. Inf. Forensics
tor-based recognition algorithms, such as local binary Secur. 11(7), 1630–1641 (2016)
pattern (LBP) and speeded up robust features (SURF), 8. Biswas, S.K., Milanfar, P.: One shot detection with laplacian
yield 92.95 and 94:57 recognition accuracy of cattle, object and fast matrix cosine similarity. IEEE Trans. Pattern
Anal. Mach. Intell. 38(3), 546–562 (2016)
respectively. The evaluation of experimental results on 9. Botella, G., Garcı́a, C.: Real-time motion estimation for image
5000 muzzle point image database of cattle demonstrates and video processing applications. J. Real-Time Image Process.
that real-time recognition and monitoring of cattle are 11(4), 625–631 (2016)
feasible using muzzle point images. It can provide a 10. Cangar, O., Leroy, T., Guarino, M., Vranken, E., Fallon, R.,
Lenehan, J., Mee, J., Berckmans, D.: Model-based calving
common platform for the recognition, tracking, and iden- monitor using real time image analysis. Precis. Livest. Farming 7,
tification of locomotion of postures, behavior, and health 291–298 (2007)
monitoring. The proposed approach can provide efficient 11. Cangar, Ö., Leroy, T., Guarino, M., Vranken, E., Fallon, R.,
solutions to classical animal recognition systems for animal Lenehan, J., Mee, J., Berckmans, D.: Automatic real-time mon-
itoring of locomotion and posture behaviour of pregnant cows
identification, registration, and supervision in the real prior to calving using online image analysis. Comput. Electron
[Link] future, we plan to extend the proposed approach for Agric. 64(1), 53–60 (2008)
cattle recognition in the real time. We would like to include 12. Coutinho, V.A., Cintra, R.J., Bayer, F.M., Kulasekera, S.,
the following points as part of our future work: Madanayake, A.: A multiplierless pruned dct-like transformation
for image and video compression that requires ten additions only.
1. We would like to design a robust cattle monitoring J Real-Time Image Process. 12(2), 247–255 (2016)
13. Dao, T.K., Le, T.L., Harle, D., Murray, P., Tachtatzis, C.,
system based on our proposed muzzle point recogni-
Marshall, S., Michie, C., Andonovic, I.: Automatic cattle loca-
tion algorithms that can handle large and surveillance tion tracking using image processing. In: Signal Processing
video and provide better recognition accuracy in the Conference (EUSIPCO), 2015 23rd European, pp. 2636–2640
real-time scenario. (2015)
14. Dell, A.I., Bender, J.A., Branson, K., Couzin, I.D., de Polavieja,
2. We would like to increase the performance of the
G.G., Noldus, L.P., Pérez-Escudero, A., Perona, P., Straw, A.D.,
proposed system; the muzzle point features can be Wikelski, M., et al.: Automated image-based tracking and its
fused with that of facial images of cattle for particular application in ecology. Trends Ecol. Evol. 29(7), 417–428 (2014)
biometric recognition system in real-time. 15. Diehl, C.P., Cauwenberghs, G.: Svm incremental learning,
adaptation and optimization. Proc. Int. Joint Conf. Neural Netw.
3. Finally we would like to explore the new possibility of
4, 2685–2690 (2003)
the particular machine learning approaches to extract 16. Doherr, M., Audige, L.: Monitoring and surveillance for rare
the discriminatory features from the muzzle point and health-related events: a review from the veterinary perspective.
face image database of cattle. We postulate that it can Philos. Trans. R. Soc. Lond. B Biol. Sci. 356(1411), 1097–1106
(2001)
be helpful even better experimental results in the
17. Duyck, J., Finn, C., Hutcheon, A., Vera, P., Salas, J., Ravela, S.:
proposed multi-model based cattle recognition system Sloop: a pattern retrieval engine for individual animal identifi-
in the real-time scenario. cation. Pattern Recognit. 48(4), 1059–1073 (2015)
18. El-Henawy, I., El Bakry, H.M., El Hadad, H.M.: Cattle identifi-
cation using segmentation-based fractal texture analysis and
artificial neural networks. Int. J. Electron. Inf. Eng. 4(2), 82–93
(2016)
References 19. Etemad, K., Chellappa, R.: Discriminant analysis for recognition
of human face images. JOSA A 14(8), 1724–1733 (1997)
1. Ahonen, T., Hadid, A., Pietikainen, M.: Face description with 20. Finn, C., Duyck, J., Hutcheon, A., Vera, P., Salas, J., Ravela, S.:
local binary patterns: application to face recognition. IEEE Trans. Relevance feedback in biometric retrieval of animal photographs.
Pattern Anal. Mach. Intell. 28(12), 2037–2041 (2006) In: Pattern Recognition, Springer, pp. 281–290 (2014)
2. Ariff, M., Ismarani, I., Shamsuddin, N.: Rfid based systematic 21. Gu, Q., Aoyama, T., Takaki, T., Ishii, I.: High frame-rate tracking
livestock health management system. In: IEEE Conference on of multiple color-patterned objects. J. Real-Time Image Process.
Systems, Process and Control (ICSPC), pp. 111–116 (2014) 11(2), 251–269 (2016)
3. Awad, A.I.: From classical methods to animal biometrics: a 22. Hadad, H.M.E., Mahmoud, H.A., Mousa, F.A.: Bovines muzzle
review on cattle identification and tracking. Comput. Electron classification based on machine learning techniques. Procedia
Agric. 123, 423–435 (2016) Comput. Sci. 65, 864–871 (2015)
4. Bartlett, M.S., Movellan, J.R., Sejnowski, T.J.: Face recognition 23. He, X., Zhang, C., Zhang, L., Li, X.: A-optimal projection for
by independent component analysis. IEEE Trans. Neural Netw. image representation. IEEE Trans. Pattern Anal. Mach. Intell.
13(6), 1450–1464 (2002) 38(5), 1009–1015 (2016)

123
 Author's personal copy
J Real-Time Image Proc

24. Hoy, J., Koehler, P., Patterson, R.: A microcomputer-based sys- 42. Lv, Z., Chirivella, J., Gagliardo, P.: Bigdata oriented multimedia
tem for real-time analysis of animal movement. J. Neurosci. mobile health applications. J. Med. Syst. 40(5), 1–10 (2016a).
Methods 64(2), 157–161 (1996) doi:10.1007/s10916-016-0475-8
25. Huhtala, A., Suhonen, K., Mäkelä, P., Hakojärvi, M., Ahokas, J.: 43. Lv, Z., Li, X., Zhang, B., Wang, W., Zhu, Y., Hu, J., Feng, S.:
Evaluation of instrumentation for cow positioning and tracking Managing big city information based on webvrgis. IEEE Access.
indoors. Biosyst. Eng. 96(3), 399–405 (2007) 4, 407–415 (2016b)
26. Huircán, J.I., Muñoz, C., Young, H., Von Dossow, L., Bustos, J., 44. Martiskainen, P., Järvinen, M., Skön, J.P., Tiirikainen, J.,
Vivallo, G., Toneatti, M.: Zigbee-based wireless sensor network Kolehmainen, M., Mononen, J.: Cow behaviour pattern recog-
localization for cattle monitoring in grazing fields. Comput. nition using a three-dimensional accelerometer and support vec-
Electron. Agric. 74(2), 258–264 (2010) tor machines. Appl. Anim. Behav. Sci. 119(1), 32–38 (2009)
27. Jain, A., Hong, L., Bolle, R.: On-line fingerprint verification. 45. Mörwald, T., Prankl, J., Zillich, M., Vincze, M.: Advances in
IEEE Trans. Pattern Anal. Mach. Intell. 19(4), 302–314 (1997) real-time object tracking. J. Real-Time Image Process. 10(4),
28. Jegadeesan, S., Venkatesan, G.P.: Smart cow health monitoring, 683–697 (2015)
farm environmental monitoring and control system using wireless 46. Ni, L.M., Liu, Y., Lau, Y.C., Patil, A.P.: Landmarc: indoor
sensor networks. Int. J. Adv. Eng. Tech./Vol VII/Issue I/Jan– location sensing using active rfid. Wirel. Netw. 10(6), 701–710
March 334, 339 (2016) (2004)
29. Johnston, A., Edwards, D.: Welfare implications of identification 47. Nilsson, M., Herlin, A., Ardö, H., Guzhva, O., Åström, K.,
of cattle by ear tags. Vet. Rec. 138(25), 612–614 (1996) Bergsten, C.: Development of automatic surveillance of animal
30. Kim, T.K., Wong, S.F., Stenger, B., Kittler, J., Cipolla, R.: behaviour and welfare using image analysis and machine learned
Incremental linear discriminant analysis using sufficient spanning segmentation technique. Animal 9(11), 1859–1865 (2015)
set approximations. In: IEEE Conference on Computer Vision 48. Oczak, M., Ismayilova, G., Costa, A., Viazzi, S., Sonoda, L.T.,
and Pattern Recognition (CVPR’07), pp. 1–8 (2007) Fels, M., Bahr, C., Hartung, J., Guarino, M., Berckmans, D.,
31. Kumar, S., Singh, S.K.: Feature selection and recognition of et al.: Analysis of aggressive behaviours of pigs by automatic
muzzle point image pattern of cattle by using hybrid chaos bfo video recordings. Comput. Electron. Agric. 99, 209–217 (2013)
and pso algorithms. In: Advances in Chaos Theory and Intelligent 49. Ojala, T., Pietikäinen, M., Mäenpää, T.: Multiresolution gray-
Control, Springer, pp. 719–751 (2016) scale and rotation invariant texture classification with local binary
32. Kumar, S., Tiwari, S., Singh, S.K.: Face recognition for cattle. In: patterns. IEEE Trans. Pattern Anal. Mach. Intell. 24(7), 971–987
2015 Third International Conference on Image Information Pro- (2002)
cessing (ICIIP), pp. 65–72 (2015a). doi:10.1109/ICIIP.2015. 50. Pisano, E.D., Zong, S., Hemminger, B.M., DeLuca, M., Johnston,
7414742 R.E., Muller, K., Braeuning, M.P., Pizer, S.M.: Contrast limited
33. Kumar, S., Tiwari, S., Singh, S.K.: Face recognition for cattle. In: adaptive histogram equalization image processing to improve the
3rd IEEE International Conference on Image Information Pro- detection of simulated spiculations in dense mammograms.
cessing (ICIIP), pp. 65–72 (2015b) J. Digit. Imaging 11(4), 193–200 (1998)
34. Kumar, S., Singh, S.K., Dutta, T., Gupta, H.P.: Poster: a real-time 51. Poiesi, F., Cavallaro, A.: Predicting and recognizing human
cattle recognition system using wireless multimedia networks. In: interactions in public spaces. J. Real-Time Image Process. 10(4),
Proceedings of the 14th Annual International Conference on 785–803 (2015)
Mobile Systems, Applications, and Services Companion, ACM, 52. Rodrı́guez-Prieto, V., Vicente-Rubiano, M., Sanchez-Matamoros,
New York, NY, USA, MobiSys ’16 Companion, pp. 48–48 A., Rubio-Guerri, C., Melero, M., Martinez-Lopez, B., Martinez-
(2016a). doi:10.1145/2938559.2948871 Aviles, M., Hoinville, L., Vergne, T., Comin, A., et al.: Sys-
35. Kumar, S., Tiwari, S., Singh, S.K.: Face recognition of cattle: tematic review of surveillance systems and methods for early
Can it be done? Proc. Natl. Acad. Sci., India Sect. A Phys. Sci. detection of exotic, new and re-emerging diseases in animal
86(2), 137–148 (2016b) populations. Epidemiol. Infect. 143(10), 2018–2042 (2015)
36. Laadjel, M., Bouridane, A., Nibouche, O., Kurugollu, F., Al- 53. Saeidi, R., Astudillo, R.F., Kolossa, D.: Uncertain LDA:
Maadeed, S.: An improved palmprint recognition system using including observation uncertainties in discriminative transforms.
iris features. J. Real-Time Image Process. 8(3), 253–263 (2013) IEEE Trans. Pattern Anal. Mach. Intell. 38(7), 1479–1488 (2016)
37. Leroy, T., Vranken, E., Van Brecht, A., Struelens, E., Janssen, A., 54. Seo, H.J., Milanfar, P.: Face verification using the lark repre-
Tuyttens, F., De Baere, K., Zoons, J., Sonck, B., Berckmans, D.: sentation. IEEE Trans. Inf. Forensics Secur. 6(4), 1275–1286
A quantitative computer vision method for on-line classification (2011)
of poultry behavior in furnished cages. Trans. ASAE 49(3), 55. Shao, J., Xin, H., Harmon, J.: Comparison of image feature
795–802 (2005) extraction for classification of swine thermal comfort behavior.
38. Lind, N.M., Vinther, M., Hemmingsen, R.P., Hansen, A.K.: Comput. Electron. Agric. 19(3), 223–232 (1998)
Validation of a digital video tracking system for recording pig 56. Tan, K., Wasif, A., Tan, C.: Objects tracking utilizing square grid
locomotor behaviour. J. Neurosci. Methods 143(2), 123–132 rfid reader antenna network. J. Electromagn. Waves Appl. 22(1),
(2005) 27–38 (2008)
39. Liu, C., Wechsler, H.: Comparative assessment of independent 57. Tillett, R., Onyango, C., Marchant, J.: Using model-based image
component analysis (ica) for face recognition. In: International processing to track animal movements. Comput. Electron. Agric.
Conference on Audio and Video Based Biometric Person 17(2), 249–261 (1997)
Authentication, Citeseer, pp. 22–24 (1999) 58. Turk, M.A., Pentland, A.P.: Face recognition using eigenfaces.
40. Lowe, D.G.: Object recognition from local scale-invariant fea- In: IEEE Computer Society Conference on Computer Vision and
tures. In: Proceedings of the Seventh IEEE International Con- Pattern Recognition (CVPR ’91), pp. 586–591 (1991)
ference on Computer Vision, vol. 2, pp. 1150–1157 (1999) 59. Wardrope, D.: Problems with the use of ear tags in cattle. Vet.
41. Lv, Z., Tek, A., Da Silva, F., Empereur-mot, C., Chavent, M., Rec. (United Kingdom) 37(26), 675 (1995)
Baaden, M.: Game on, science—how video game technology 60. Weng, J., Zhang, Y., Hwang, W.S.: Candid covariance-free
may help biologists tackle visualization challenges. PLoS One incremental principal component analysis. IEEE Trans. Pattern
8(3), 1–13 (2013). doi:10.1371/[Link].0057990 Anal. Mach. Intell. 25(8), 1034–1040 (2003)

123
 Author's personal copy
J Real-Time Image Proc

61. Wolf, L., Hassner, T., Taigman, Y.: The one-shot similarity Pattern Recognition, and Artificial Intelligence. He has over 50
kernel. In: 12th IEEE International Conference on Computer national and international journal publications, book chapters, and
Vision, pp. 897–902 (2009) conference papers. He is also a Guest Editorial Board Member of
62. Wolf, L., Hassner, T., Taigman, Y.: Effective unconstrained face Multimedia Application and Tools, Springer, and the EURASIP
recognition by combining multiple descriptors and learned Journal of Image and Vision Processing (Springer). He is a member of
background statistics. IEEE Trans. Pattern Anal. Mach. Intell. the Computer Society and the Association for Computing Machinery.
33(10), 1978–1990 (2011)
63. Yang, J., Lin, Y., Gao, Z., Lv, Z., Wei, W., Song, H.: Quality Ravi Shankar Singh is an Assistant Professor in the Department of
index for stereoscopic images by separately evaluating adding Computer Science and Engineering, (B.H.U.), Varanasi, India, since
and subtracting. PLoS One 10(12), 1–19 (2016). doi:10.1371/ 2004. His research interests include tasks scheduling techniques in
[Link].0145800 distributed computing, data structures, parallel algorithms. He is
64. Yang, L., Jin, R.: Distance Metric Learning: A Comprehensive member of ACM and Computer Society of India. He has presented
Survey. Michigan State Universiy, vol. 2 (2006) and published over30 research papers in various national and
65. Zhu, Q., Ren, J., Barclay, D., McCormack, S., Thomson, W.: international conferences.
Automatic animal detection from kinect sensed images for live-
stock monitoring and assessment. In: IEEE International Con- Amit Kumar Singh is currently working as Assistant Professor in the
ference on Computer and Information Technology; Ubiquitous Department of Computer Science and Engineering at Jaypee Univer-
Computing and Communications; Dependable, Autonomic and sity of Information Technology (JUIT) Waknaghat, Solan, Himachal
Secure Computing; Pervasive Intelligence and Computing (CIT/ Pradesh, India, since April 2008. He has completed his PhD degree
IUCC/DASC/PICOM), pp. 1154–1157 (2015) from the Department of Computer Engineering, NIT Kurukshetra,
Haryana, in 2015. He has presented and published over 40 research
Santosh Kumar received the B. Tech. degree in Computer Science
papers in reputed journals and various national and international
and Engineering, from the Uttar Pradesh Technical University
conferences. His important research contributions includes to develop
(UPTU), in 2008, and M. Tech. in Computer Science and Engineering
watermarking methods that offer a good trade-off between major
from Birla Institute of Technology (B.I.T), Mesra, Ranchi (Jharkhand,
parameters, i.e., perceptual quality, robustness, embedding capacity,
India) in 2012. Currently he is pursuing PhD from the Department of
and the security of the watermark embedding into the cover digital
Computer Science and Engineering, Indian Institute of Technology
images. Recently, he has appointed as member of editorial board in
(B.H.U.), Varanasi-221005. His research interests span a wide range
one of the prestigious image processing journal Multimedia tools and
of spectrum including Computer vision, Pattern Recognition, Digital
Applications (MTAP). Currently, Dr. Singh has served as Guest
Image Processing, Digital Video Processing, Swarm Intelligence,
editor, TPC member, and reviewer for various conferences and
Metaheuristics, Bio-inspired Computing, Artificial Intelligence. He
reputed journals. His research interests include Data Hiding,
was a recipient of best poster awards at Institute Day, Indian Institute
Biometrics and Cryptography.
of Technology (B.H.U.), Varanasi-221005. He is a member of the
Computer Society and the Association for Computing Machinery.
Shrikant Tiwari received his PhD degree from the Department of
Computer Science and Engineering, Indian Institute of Technology
Sanjay Kumar Singh has completed his B. Tech. in Computer Engg.,
(B.H.U), India, in 2012. He is an Assistant Professor in the
M. Tech. in Computer Applications, and PhD in Computer Science
Department of Computer Science and Engineering, Shri Shankar-
and Engineering. Currently he is an Associated Professor at the
acharya Technical Campus, Junwani, Bhilai, District-Durg, Chattis-
Department of Computer Science and Engineering, IIT BHU,
grah, India-490020. His research interests include Image Processing,
Varanasi. He is a Certified Novell Engineer (CNE) from Novell
Machine Learning, and their applications in biometrics. He has
Netware, USA, and a Certified Novell Administrator (CNA) from
presented and published over 20 research papers and book chapters in
Novell Netware, USA. He is a member of LIMSTE, IEE, Interna-
various national and international conferences.
tional Association of Engineers and ISCE. His research areas include
Biometrics, Computer Vision, Image Processing, Video Processing,

123

View publication stats