See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/320686855

Deep Learning Framework for Recognition of Cattle using Muzzle Point Image
Pattern

Article  in  Measurement · October 2017

DOI: 10.1016/j.measurement.2017.10.064

CITATIONS READS

21 3,967

7 authors, including:

Santosh Kumar Amit Pandey

IIIT Naya Raipur, Chhattisgarh Indian Institute of Technology (Banaras Hindu University) Varanasi
55 PUBLICATIONS   378 CITATIONS    1 PUBLICATION   21 CITATIONS   

SEE PROFILE SEE PROFILE

Satwik Kondamudi Sunil Kumar

Banaras Hindu University Shri Ramswaroop Memorial University
1 PUBLICATION   21 CITATIONS    33 PUBLICATIONS   457 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Machine Learning Paradigms: Theory and Applications Security in Smart Cities View project

Call for papers - Cognitive Big Data Analytics for Business Intelligence applications: Towards Performance Improvement, International Journal of Information
Management, Elsevier (SCI Indexed, IF = 4.516) View project

All content following this page was uploaded by Santosh Kumar on 07 November 2017.

The user has requested enhancement of the downloaded file.

 Measurement 116 (2018) 1–17

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Deep learning framework for recognition of cattle using muzzle point image MARK
pattern
⁎
Santosh Kumara, , Amit Pandeyb, K. Sai Ram Satwikb, Sunil Kumarc, Sanjay Kumar Singhd,
Amit Kumar Singhe, Anand Mohanf
a
Computer Science and Engineering, Dr. Shyama Prasad Mukhurjee, IIIT Naya Raipur, Chhattisgarh 493661, India
b
Department of Electrical Engineering, Indian Institute of Technology (B. H. U.), Varanasi, Uttar Pradesh 221005, India
c
Indian Maritime University – Kolkata Campus, 700088, India
d
Department of Computer Science and Engineering, Indian Institute of Technology (BHU), Varanasi, Uttar Pradesh 221005, India
e
Department of Computer Science and Engineering, Jaypee University of Information Technology, Solan, Himachal Pradesh 173234, India
f
Department of Electronics Engineering, Indian Institute of Technology (BHU), Varanasi, Uttar Pradesh 221005, India

A R T I C L E I N F O A B S T R A C T

Keywords: Animal biometrics is a frontier area of computer vision, pattern recognition and cognitive science to plays the
Cattle recognition vital role for the registration, unique identification, and verification of livestock (cattle). The existing hand-
Muzzle point image crafted texture feature extraction and appearance based feature representation techniques are unable to perform
Deep learning the animal recognition in the unconstrained environment. Recently deep learning approaches have achieved
Convolution Neural Network
more attention for recognition of species or individual animal using visual features. In this research, we propose
DBN
the deep learning based approach for identification of individual cattle based on their primary muzzle point
SDAE
Verification (nose pattern) image pattern characteristics to addressing the problem of missed or swapped animals and false
Computer vision insurance claims. The major contributions of the work as follows: (1) preparation of muzzle point image da-
LBP tabase, which are not publically available, (2) extraction of the salient set of texture features and representation
SURF of muzzle point image of cattle using the deep learning based convolution neural network, deep belief neural
PCA network proposed approaches. The stacked denoising auto-encoder technique is applied to encode the extracted
VLAD feature of muzzle point images and (3) experimental results and analysis of proposed approach. Extensive ex-
LDA
perimental results illustrate that the proposed deep learning approach outperforms state-of-the-art methods for
recognition of cattle on muzzle point image database. The efficacy of the proposed deep learning approach is
computed under different identification settings. With multiple test galleries, rank-1 identification accuracy of
98.99% is achieved.

1. Introduction biometric features from the morphological image, biometric char-

acteristics and phenotypic appearances of different species or in-
Animal biometrics is an emerging research field of computer vision, dividual animal.
wildlife science, and pattern recognition [1]. Animal biometrics-based In animal biometrics, identification of cattle based on biometric
recognition system develops quantified and efficient recognition features has been one of the current and future research frontiers in the
methodologies for representing extracted visual features, detecting modern livestock for registration, tracking and breed associations of
discriminatory features for identifying the phenotypic appearance of cattle.
species or analysis of individual animal’s behaviours based on its Based on available literature, animal identification methodologies
morphological image pattern and animal biometric characteristics. The can be categorized into following groups: (1) permanent recognition
phenotypic presentations consist of the composite of an organism's method, (2) semi-permanent recognition method, and (3) sketch pat-
observable morphological features [1,2]. tern-based marking recognition approach [1,2,5]. The permanent re-
Animal biometrics-based recognition system is a pattern recognition cognition method includes the making based animal identification
system. The recognition system extracts the prominent animal schemes. These marking schemes are as ear-tattoo-based identification

⁎
Corresponding author.
E-mail addresses: [email protected] (S. Kumar), [email protected] (A. Pandey), [email protected] (K. Sai Ram Satwik),
[email protected] (S. Kumar), [email protected] (S.K. Singh), [email protected] (A.K. Singh), [email protected] (A. Mohan).

http://dx.doi.org/10.1016/j.measurement.2017.10.064
Received 21 November 2016; Received in revised form 17 September 2017; Accepted 27 October 2017
Available online 28 October 2017
0263-2241/ © 2017 Elsevier Ltd. All rights reserved.
S. Kumar et al. Measurement 116 (2018) 1–17

techniques, embedded of microchips-based marking approach invasive, ear-tags use low-frequency (LF) RFID. Therefore, the RFID scanner
hot-iron and freeze branding of livestock. The ear-tattoos-based animal systems must be deployed within a few inches of the labelled ear-tags.
marking schemes are the manual marking technique to identify the The users (e.g., farm workers, parentages, owners of cattle and others)
massive number of livestock animal’s herds and other species. The require operating the scanned devices by personal with every animal
sketch patterning-based marker methods are not used to identify the for reliable scanning of embedded ear-tags.
solid colour breeds of livestock (especially cattle breeds). The RFID based ear-tagging of individual cattle has been widely
The ear-tattoos-based marking techniques are highly applicable for adopted throughout the world. For example, dairy farms in the Europe,
tracking and identification of several cattle breed associations such as USA, and Asian countries are more reluctant to perform the identifi-
Brown Swiss, Red Poll, and Milking Shorthorn breed associations [5,6]. cation of cattle so due to costs of the ear-tags, and the lack of national
The distinct artificial marking techniques, such as embedding of mi- standards.
crochips, ear-tattooing and hot-iron techniques are required for iden- In the classical animal identification approaches, animal recognition
tifying individual cattle, but these techniques give the defects on the and livestock framework-based systems provide manual procedures or
cattle [6,7]. methods for verifying the registered (e.g., insurance animal) and im-
On the other hand, the collar-id and ear-tagging based identification poster animal based on registered unique ear tag numbers of livestock,
techniques are the examples of semi-permanent animal identification embedded number in hot iron, and marked number in the ear-tattoos on
approaches for identification of individual animals [8,9]. According to body surface of animal. The embedded unique numbers can be easily
authors Wardrope [6], Johnston and Edwards [10], the ear-tagging duplicated, and forged for creating the false registered number. The
based verification and identification techniques are not competent to users (parentage, owners and others) claim number of false insurance
identify individual cattle. The monitoring and tracking of individual claims by replacing the previous registered unique number. In the
cattle based on embedded unique tag numbers can be easily lost. available literature, the classical animal recognition and livestock fra-
The ear-tagging based techniques suffer from the major problems mework based systems cater the manual methods for verifying the re-
for verification and identification of animals. These major problems of gistered (e.g., insurance animals by government and insurance provider
ear-tagging based techniques are mainly-(1) integrated labels in the ear- organizations) to find the imposter animals. These recognition systems
tags can vanish easily (2) different integrated labels can also be even- use the manual approaches such as ear tags, hot iron, and ear tattoos.
tually damaged and (3) the ear can gradually be corrupted due to long- Therefore, classical identification methods are unable to provide the
term usage [6–10]. satisfactory level of security for owners and verification of cattle breed
The sketch based patterning techniques perform the design and associations in the farmhouses throughout the world.
different patterning of different colour on the body of livestock. The Animal biometrics-based cattle recognition system gives the effi-
sketch patterning and design-based techniques depend on individual cient methodologies for registration of livestock animals using its bio-
drawing skills. The major problem of patterning technique is that it metric features. The cattle recognition system also caters the better
lacks standards [7]. Due to the enormous amount of human resource method for monitoring, and traceability of individual cattle in the
requirements, the classical animal identification approaches need more farmhouses [2,4]. Moreover, the recognition system plays a vital role in
cost for the tagging of cattle’s ear. The ear-tagging based systems suffers the verification of false insurance claims and registration of livestock,
from the venerability loses due to duplication, fraudulent and forged of tracking, breeding monitoring, and breeds production. Nowadays, non-
embedded standard ear-tags [2,3,5,12]. The embedded RFID chips registration and the massive number of livestock animals are trans-
based techniques are also applied for the identification and tracking of ported across the border, and cattle recognition system provides effi-
individual cattle. However, the implementation and management of cient methods to stop the kind of activities of border transfer of live-
RFID protocol, RFID-chips, and scanners at various checkpoints have stock animals for slaughterhouses [5,6].
been reported the major challenging problems for monitoring of the Currently, deep learning approach is an emerging field of computer
massive number of livestock animal and other species throughout the vision for detection and representation of phenotypic appearances and
world [12]. visual biometric feature of species and individual animals. Deep
learning has gained more attention as of the most powerful approaches
1.1. Motivation for the feature extraction and representation of different species or in-
dividual animal in the recent years [7–9]. With efficient learning ca-
Towards successful operation of any farm, an effective livestock pacity for feature representations, the deep learning based learning
management is essential. Efficiency, affordability, and scalability of li- frameworks are used to learn the extracted sets of biometric feature of
vestock management and solutions play a significant role in the modern species for the representations and identification [10,11]. The different
farm houses. As the number of farms decreases in the country, but the layers of the deep learning framework are used to model and represent
numbers of livestock on each increases every year. Dairy livestock the complex data variations for animal recognition [12,13].
needs better monitoring for breed associations, milking, health mon- The well-developed deep learning approaches and frameworks are
itoring of animals, weighting and other activities. Therefore, the ability applied to propose a cattle recognition system to solve the above issues
to reliably tracking and identifying individual animal in herds is an by learning the muzzle point feature its representation and classifiers
important for cattle registration and recognition. jointly for a particular task tasks [14,15].
The cattle recognition using manual identification approaches has In this paper, we address the problem: how to recognize cattle based
been a major problem for breeding association, registration and health biometric feature characteristics of muzzle point image pattern (nose
monitoring of the livestock (especially for cattle) in the classical animal image pattern) using deep learning-based recognition approaches? To
identification systems, livestock based monitoring frameworks, and address this problem, we use a 30 megapixel camera for capturing the
non-biometrics-based capture- recapture marking methodologies [3,4]. image of muzzle point pattern of cattle.
The classical animal identification framework and identification The muzzle point image pattern consists of rich and dense texture
paradigm provides non-invasive technique for identification of in- features. The proposed cattle recognition system recognize the cattle
dividual cattle in the heard. The non-invasive identification techniques based on texture feature of muzzle point image. The system uses the
such as visible ear tagging based identification, freeze brandings, deep learning based recognition frameworks.
marking the animal body with hot-iron, embedding of microchips are The cattle recognition system captures the image of muzzle point
applied for cattle registration. Moreover, the labelled ear-tags with pattern (nose pattern) of cattle. After that, the captured image of
RFID devices have been also incorporated for cattle identification. The muzzle point pattern are pre-processed using the low pass filtering
labelled ear-tags are scanned electronically because of cost, and most technique to remove the noises and other artefacts from the captured

2
S. Kumar et al. Measurement 116 (2018) 1–17

muzzle point image database. The texture features are extracted from approaches. Stacked Denoising Auto-encoder is presented in Section 5.
the muzzle point image database. The extracted features are classified Section 6 provides the Pre-Training and generalizability of proposed
using deep learning approaches to identify the individual cattle. model in brief. The experimental results and brief analysis of the work
Based on overall observations and our acknowledgment, there are is reported in Section 7. Further, it includes the performance evalua-
no such animal biometrics-based recognition and verification systems tions and comparative analysis of proposed approach. Finally, the
present in animal biometrics literature or public domain to cater the conclusions and future directions are provided in Section 8.
better methods for identification and registration of cattle based on
primary biometric characteristics. 2. Related work

1.2. Research contributions In this section, we have provided the literature review of cattle
using various computer vision and animal biometrics approaches for
Following are important contributions of this research: identification of individual cattle. Off late the animal biometrics-based
recognition systems have received significant attentions due to more
1. The unique and immutable bead and ridge pattern as biometric proliferation for recognition, detection, tracking, and monitoring of
characteristics of muzzle point image of cattle leads to interesting different species or individual animal [1,2]. These animal biometrics-
challenges for an animal biometrics-based recognition system for based recognition systems identify the individual animal or species
cattle identification. Considering the non-intrusive nature of muzzle using its physical visual features (e.g., facial images), morphological
point biometrics pattern, this research explores the new possibility image pattern, and biometric characteristics of animals. The permanent
of determining the unique identity of individual cattle using deep animal recognition techniques provide the safety by recognizing the
learning approaches. Motivated from animal biometrics research animal using ear-tattoos, ear-tip or notches, freeze-branding, hot iron
across multiple domains and interdisciplinary researches, we discuss around animal’s neck and embedded microchips, electronic devices,
the major challenges and opportunities of the cattle recognition sensors, and RFID-transponders [3–5].
system. For recognition of individual cattle, novel deep learning For example, ear tagging, and ID-collar-based recognition techni-
based representation frameworks and matching schemes are pro- ques are semi-permanent based recognition approaches for the identi-
posed in this paper. It is customized towards recognition of cattle fication of individual cattle [5,6]. However, the embedded label of ear-
and yields current state-of-the-art results. tags can be lost quick. The ear-tagging systems include only the various
2. In this paper, a novel deep learning-based stacked denoising auto- kinds of metal clips and plastic labels or tags. However, these ear-tags
encoder framework is used to encode and decode the extracted set of can cause different types of infections and critical diseases to animals
salient texture feature of muzzle point image of cattle for recogni- after these embedding of these tags in the body of the animal or other
tion of individual cattle. species.
3. The deep learning framework-based approaches are suitable to ad- The ear-tagging and collar-ID based approaches are susceptible to
dress the significant variations of muzzle point images of cattle in damage, duplication, losses, un-readability, and fraud of ear-tags.
the unconstrained conditions due to low illumination, body dy- Therefore these methods are not fit for long-term usages [6]. While in
namics and blurred images due to the head movement and body temporary recognition based methodologies, it recognizes the in-
dynamics of cattle, and the poor image quality of muzzle point dividual animal by utilizing the sketch patterning-based techniques. For
images. We proposed a Convolution Neural Network (CNN) and example, paint or dying and RFID based animal recognition with em-
Deep Belief Network (DBN) deep learning approaches to learn the bedded transponders or sensors in their body are temporary identifi-
muzzle point texture feature of cattle for better representation into cation methods [7–11].
different deep neural network layers with deep learning framework. Based on the available classical identification techniques, the RFID-
4. A deep belief network (DBN) is the graphical deep learning model is based animal identification process is one of the most promising for
used to learn the extracted features of muzzle images for hier- cattle or other livestock animals. According to the author Baranov et al.
archical representation of the training datasets by stacking of [16], the muzzle dermatoglyphics (i.e., ridges, granola, and vibrissae)
Restricted Boltzmann Machines (RBMs) based classification tech- from various races are different. In a similar direction, Mishra et al.
nique for classification and identification of cattle. It is trained in a [17] proposed a method for identification of individual cattle using
greedy manner for learning of feature extraction and representation beads and ridges pattern.
from the given large unlabelled image database and updating the Minagawa et al. [18] introduced a framework for the cattle identi-
loss function of DBM by including low rank regularization. Finally, a fication using muzzle print images. The muzzle print images were
multilayer deep learning neural network has been used as classifier captured on A-5 paper with black ink and evaluated performances of
to achieve the identification decision for cattle. proposed approach using filtering techniques. For analysis of muzzle
5. A database of muzzle point image of cattle over 5000 images per- print image, the authors applied the binary transformation processes
taining to over 500cattle is contributed to the research community and morphological approaches (i.e., thinning operation). They have
in the animal biometrics. To the best of our knowledge, this is the reported the Equal Error Rate (EER) of 0.419, respectively.
largest muzzle point image dataset publicly available for research on Barry et al. [19] proposed a technique for recognition of cattle using
cattle recognition. Moreover, the performances of existing hand- muzzle print image similar to Minagawa et al. [18]. They have done
crafted texture feature extraction technique and appearance-based experimentation to evaluate the performance of proposed approach.
feature extraction and representation algorithms systems are com- They have reported the 241 False Non-Match Rates (F NMR) over 560
pared and evaluated the experimental results across various gallery Genuine Acceptance Rate (GAR), and 5197 false matches over 12,160
sizes of muzzle point images on a standard identification settings impostors matching closely with the same value of EER of 0.429, re-
and benchmark. spectively.
Awad et al. [20] proposed a cattle recognition framework using
The remaining parts of the paper are as follows: Section 2 illustrates SIFT descriptor approach to localize and detect the interesting points in
the literature review in the direction of cattle recognition. Section 3 the muzzle print images for the identification of cattle database is 90
provides the database preparation and description of muzzle point muzzle images (6 × 15 = 90).
images of cattle. Section 4 detailed the proposed deep learning based In the similar direction, the author of [21] proposed a matching
recognition system of cattle followed by the feature extraction and re- refinement technique in Scale Invariant Feature Transform (SIFT) de-
presentation of muzzle point image features using deep learning scriptor approach for the recognition of cattle on the database of 160

3
S. Kumar et al. Measurement 116 (2018) 1–17

muzzle print images. They have computed the matching scores of the muzzle point image pattern of cattle is shown in Fig. 1.
key points of muzzle impact images by applying matching refinement The database contains images of muzzle point pattern in the form of
technique in SIFT key point localization and detection approach. various covariates, such as low illumination, poor image quality, pose
However, the matching refinement approach has had the performance variation due to body dynamics and head movement images (shown in
compared to the original SIFT approach with the value of EER being Fig. 2). From these muzzle point images, we manually filtered the
equal to 0.0167 [23]. covariates images along with blurred and low illumination muzzle
In the same direction, Tillett et al. [22] used the image processing point images [23]. In total, muzzle point image pattern database of
and computer vision based techniques for the recognition and analysis cattle, therefore, the database consists of 500 muzzle point images
of the behaviour of pigs. Pigs are identified in surveillance video by about 500 subjects. The extraction of discriminatory texture muzzle
tracking and detecting their movements. point features (bead and ridge pattern is shown in Fig. 2.
Kumar et al. [23] proposed a framework to determine the individual
cattle identity based on their face image. They have evaluated the
performance proposed system using extracted features. 3.1. Biometric characteristics of muzzle point image pattern of cattle
Recently, author Tharwat et al. [24] proposed a muzzle image based
cattle recognition approach using local texture descriptor based tech- For identification and traceability of animals, smart devices (e.g.,
nique such as LBP texture algorithm for the extraction of local texture smart phone, digital camera, smart watches, and distributed embedded
features from the muzzle print images. The RANdom SAmple Consensus smart camera) are gaining diverse proliferation due to the wide range of
(RANSAC) approach is utilized to mitigating the outliers from the applications and uses. It is becoming a good platform for sensing the
muzzle image which is incorporated in the SIFT algorithm for the im- multimedia data, computing, and communication. For identification of
provement of reliability and robustness of their proposed identification individual cattle using proposed cattle recognition system, a 30-mega-
approach for cattle. The six images for each subject (cattle) are used as pixel camera is used to capture the muzzle point image of cattle.
training muzzle images in the experiment, and the total size of the The recognition of animals is very similar to the identification of
database is 90 muzzle images (6 × 15 = 90). They achieved more than minutiae points in the human fingerprint. The motivation of this the
90% accuracy of cattle recognition. The major shortcoming of this research works for cattle recognition is that the muzzle point image
proposed approach is to take more processing time during recognition pattern of cattle consists of rich dense texture feature. The muzzle point
of livestock (cattle). image pattern of cattle consists of two discriminatory features-(1) beads
image feature pattern (2) ridge image feature pattern of muzzle point
image of cattle. The bead image features are the prominent set of tex-
3. Database preparation and description ture features pattern. It consists of anon-uniform image pattern in the
muzzle point images of cattle. The bead and ridge feature pattern of the
To the best of our knowledge, there is no publicly available muzzle muzzle point image of cattle are shown in Fig. 3 and Fig. 4, respectively.
point image pattern database that can be applied to evaluate the current The ridges features are the uniform image pattern. It is similar to
state-of-the-art based recognition and classification algorithms or de- ridges of the human fingerprint images. The texture feature contains the
velop new algorithms for recognizing and verifying the muzzle point discriminatory patterns as bead pattern and ridge pattern for cattle
image pattern of cattle. Inspired from the classical animal recognition recognition. The recognition of cattle based on muzzle point image is
systems and its failures, there is a need to develop an automatic, non- similar to recognition of human fingerprint recognition. Therefore, in
invasive and robust animal biometric systems for identification of this research, we proposed a novel approach for recognition of cattle
cattle, therefore, we prepare a database of muzzle point image pattern based on muzzle point image characteristics which start with selecting
of cattle using a 30-megapixel camera from the Department of Dairy texture rich feature of the muzzle point image database of cattle using
and Husbandry, Institute of Agricultural Sciences (I. A. S), Banaras deep learning convolutional neural network (CNN) architecture which
Hindu University (B.H.U), Varanasi, India-221005. Some sample of is a combination of Stacked Denoising Sparse Auto-encoder (SDSA) and

Fig. 1. Some muzzle point images of cattle from database.

4
S. Kumar et al. Measurement 116 (2018) 1–17

Fig. 2. Illustrates the blurred muzzle point images in (1, 2, 3 and 6) and images (4 and 5) shows poor illumination.

Deep Boltzmann machine (DBM) learning techniques. of muzzle point of cattle which are captured in the unconstrained en-
vironment. Therefore, texture feature descriptor techniques are utilized
4. Proposed system for uniquely identification of individual cattle in this work.
The proposed cattle recognition system is used to learn the dis-
A deep learning based framework is used in the proposed cattle criminatory set of extracted muzzle point image features for better re-
recognition system for identification of cattle is depicted in Fig. 6. presentation with limited training database. The basis of the proposed
The proposed cattle recognition system takes the captured muzzle approach includes muzzle point texture feature pattern of cattle for
point image pattern of cattle. The captured images of muzzle point recognition of cattle. The muzzle point features consist of two promi-
pattern consist of rich dense texture feature of muzzle point image. nent sets of feature known as bead and ridge features. The features are
Therefore, the texture feature of muzzle point image are extracted and extracted and represented using proposed deep learning approaches
encoded using the deep learning frameworks and the existing hand- such as convolution neural network, deep belief network frameworks.
crafted texture feature extraction and appearance based feature re- The extracted set of salient muzzle point features of cattle are encoded
presentation techniques. The appearance based feature extraction and using stacked denosing auto-encoder technique for better feature re-
representation algorithms are unable to perform the recognition of presentation in feature space.
cattle based on low illumination, poor image quality, and blurred image The salient set of extracted muzzle point feature of cattle is

Fig. 3. Illustrates the pre-processing process (a) muzzle point image (b) filtration of discriminatory features (beads and ridges) (c) shows the removal of background (d) find out the ridges
and bead features from the overlapped features.

5
S. Kumar et al. Measurement 116 (2018) 1–17

Fig. 4. Illustrates the extraction of discriminatory feature (bead and ridge pattern) from muzzle point image.

classified by deep learning based Restricted Boltzmann Machines effective solutions for cattle registration and verification of false in-
(RBMs) classification framework is trained in a greedy manner for surance for individual cattle using the low-cost camera. It also provides
learning of feature extraction and representation from the given large the better solution for identification of cattle in the classical animal
unlabelled muzzle point image database and updating the loss function identification methodologies and classical livestock frameworks. The
of DBM by including low rank regularization. Finally, a multilayer deep proposed recognition system consists of various steps for identification
learning neural network has been used as classifier to achieve the of cattle which are depicted in the next subsection.
identification decision for cattle.
The proposed deep learning framework is applied to identify the
cattle. The primary objective of the proposed deep learning based cattle 4.1. Pre-processing and enhancement
recognition system is used for identification of individual cattle.
The purpose of animal biometrics-based cattle recognition system is In this subsection, we have applied various image pre-processing
to provide the efficient methodologies for registration of livestock an- techniques to mitigate the noises from captured muzzle point images.
imals using its biometric characteristics such as muzzle point image The fundamental problems involved in the acquisition of images of
pattern (nose image pattern) and face image biometric features. The cattle are (a) little illumination and (b) poor image quality. The capture
cattle recognition system also caters the better method for cattle muzzle point images from the unconstrained environments are trans-
monitoring, and traceability of individual cattle in the farmhouses. formed into greyscale images to mitigate the artefacts and noises from
Moreover, cattle recognition system also plays a vital role in the ver- muzzle point images [25].
ification of false insurance claims and enhance the breeding associa- After the pre-processing, the transformed images are enhanced by
tions, and breeds production. Nowadays, massive numbers of non-re- contrast limited adaptive histogram equalization (CLAHE) based image
gistered livestock animals are transferred across the borders for processing technique. It improves the contrast between patterns of
slaughterhouses. The cattle recognition system can perform the accu- muzzle point images. Moreover, CLAHE enhancement technique over
rate verification of registered and non-registered cattle to stop these amplifies the distinct noises in approximately similar regions of interest
activities of border transfer of livestock animals for slaughterhouses. in muzzle point image pattern [25].
The proposed recognition system provides automatic and cost-

Fig. 5. Illustrates the segmentation of muzzle point images: (a) illustrates the region of interest (ROI) of muzzle point images, (b) ridge bifurcation (green colour) and ridge termination
(red colour) are extracted from ROI and (c) suppression of irrelative feature information.

6
S. Kumar et al. Measurement 116 (2018) 1–17

Fig. 6. Block diagram of proposed deep learning based framework for cattle recognition.

4.2. Feature extraction and representation of muzzle point image softmax () as a non-linear activation function is chosen in the proposed
approach. The softmax () activation function provides a feature gra-
The proposed system consists of a deep learning framework for the dient at the final layer of a network used as for classification. Where(R)
learning and representations of the extracted sets of discriminatory depict a local region in the defined layers where assigned weights are
texture and holistic features of the muzzle point image pattern (as shared between the layers.
shown in Fig. 5). Since the proposed recognition system is used as the The prepared database is captured from unconstrained environ-
general in nature. It is explained using different deep learning archi- ments, such as low illumination, the poor image due to blurriness, and
tectures: Convolutional Neural Network (CNN) [26,27], Deep Belief head movement, body dynamics of cattle. To handle this problem, we
Network (DBN) [28–30] and Stacked Denoising Auto-encoders (SDAE) have applied the Convolution Neural Networks (CNNs) [8,9], for au-
[31,32]. The following subsection first explains the basics of CNN, DBN, tomatically learning of extracted discriminatory muzzle features.
and SDAE auto-encoder techniques followed by the proposed cattle In the fully-connected layers, the input muzzle images are depicted
recognition system with multiple neural networks [13]. as a vertical line of neurons. The proposed recognition system gives the
The underlying motivation behind applying deep learning based input pixels of muzzle point images to a layer of hidden neurons. For
feature extraction and representation architectures is that to learn the each local receptive field, different hidden neurons are chosen in the
discriminatory texture features of muzzle point for better representa- first hidden layer.
tion in feature space. The brief descriptions of applied deep learning After that, we start the local receptive field over by one pixel to the
methods are given in the next subsection. right (i.e., one neuron), to connect to a second hidden neuron and so on,
to building up the first hidden layer. As mentioned in the above section,
4.3. Convolution neural network each hidden neuron has a bias weight (C j ) and (5 × 5) weights con-
nected to its local receptive fields. For training, we have chosen the
Convolution Neural Network (CNN) is a deep learning approach for same (5 × 5) weight matrix and bias weight (C j ) for each of the
the feature extraction and representation through relations between the (24 × 24) hidden neurons shown as follows:
two muzzle point images patterns of cattle are modeled hierarchically. l=4 m=4
The CNN framework is built by stacking the multiple convolution layers ⎛ ⎞
σ ⎜b + ∑∑ wl,m × (aj + l,k + m)⎟
and pooling layers. ⎝ l=0 m=0 ⎠ (2)
The main motivation of the deep CNN framework-based recognition
system for cattle recognition is that CNN network extracts muzzle point In other words, for the (jth )
layers, output of (k th)
hidden neuron is given
features directly from muzzle image pixel and provided the output a in Eq. (2). Where(σ ) is the neural activation sigmoid function. (C j ) is the
highly-compact representation after training of a number of muzzle shared weight value for the bias w (l,m) . It is a (5 × 5) matrix of shared
point image pattern of cattle database. The CNN deep learning tech- weights. Finally, we use (x ) and (y ) to denote the input activation at
nique is used to accelerate the training of deep neural networks and position (x ,y ) . The first two convolution layers are followed by max-
take advantage of the multi-scale structure of the muzzle point image pooling layer for feature reduction and increasing their robustness to
pattern of cattle for cattle recognition. distortions of muzzle point image pattern.
The mathematical formulation is illustrated for feature extraction
and mapping of input muzzle point images using the convolution layer B. Max-pooling layer: The max-pooling layer is applied in the pro-
in the proposed deep learning based cattle recognition system as fol- posed approach for selection of the maximum values of every (2 × 2)
lows: grid in the muzzle point feature map. This procedure units directly
outputs the maximum activation in the input region (as shown in Eq.
A. Convolution Layer: The operation in each convolution layer is (3))
formulated as follows: P (s ) = Y ij,k = max 1 ⩽ p,q ⩽ s (x (ij − 1). s + m,(k − 1). s + n ) (3)
⏟
⎛
Y j (r ) = f ⎜C j (R) + ∑ K ij (R) ∗ (X i (R))⎞⎟ Where each neuron in the (ith) output map (Y i ) pools over a (s × s )
⎝ i ⎠ (1) non-overlapping local region in the (ith) input map (x i ) and max pooling
In the Eq. (1), (∗) denotes the convolution operation. (X i ) and (Y j ) re- layer is represented by P (s ). The final convolution layer is followed by
present the (ith) input and the (jth ) output in the CNN, respectively. The two successive fully-connected layers. The final layer of connections in
the network is a fully-connected layer. That is this layer connects every

7
S. Kumar et al. Measurement 116 (2018) 1–17

neuron from the max-pooled layer to every one of the output neurons. point image pattern of cattle and encoding the extracted set of features
for better feature representation in feature space.Stacking denoising
4.4. Classification of features using deep belief network and restricted auto-encoders is applied to initialize the deep network. It works in
Boltzmann machines much the same way as stacking the RBMs models in deep belief net-
works. The stacking of denoising auto-encoder is chosen such that the
A Deep Belief Network (DBN) is the graphical deep learning model output layer of the first auto-encoder operates as the input layer of the
which is applied to learn the extracted sets of extracted features of second auto-encoder. The two fundamental components of an auto-
muzzle point image pattern for hierarchical representation using the encoder are mainly - (1) the encoder and (2) decoder. An encoder maps
training detains the training phase of the proposed system [28,29]. DBN the input (X ) to the hidden layer nodes using deterministic mapping
learning framework is proposed by stacking of Restricted Boltzmann function (f : h = f (X )) (shown in Eqs. (4) and (5)).
Machines (RBMs) learning techniques. It is trained in a greedy manner f = GθX = s (w. X + Δ) (4)
for classification of extracted features. The main motivation to apply
the RBM a deep learning technique is that it is extremely useful for Where θ = {w,Δ} the parameter set, s is represents the sigmoid, (w) is
unsupervised learning of the feature extraction and representation of α × α weight matrix. Δ is the offset vector of size α′. Feature f is applied
muzzle point image which is taken from the given large unlabeled to map to feature vector x ′ of dimensionality α using a decoder function
muzzle point image database. G′θ such that,
For the classification, the logistic regression classifier is used based
Y ″ = G′θ′f = s (w. X ′ + Δ)
′ (5)
on h (l) (the last hidden layer of DBN learning model). This step is si-
milar with the using the weights (w) and hidden layer biases which is Where θ′ = {w′,Δ} ′ the decoder parameter is set such that argmin
generated with the unsupervised training for initialization of the (fae = ‖X −Y ″‖2F ) .
weights of a Multilayer Perceptron Layer (MLP) neural network. The A decoder maps the hidden nodes back to the original input space
training of the proposed deep DBN learning model is shown in Fig. 7. In through another deterministic mapping function (GW ′,b′). For real-va-
this approach, the DBN learning model is applied by constructing the lued input, by minimizing the reconstruction error (fae = ‖X −Y ″‖2F ) .
multiple RBM models. The RBM classification is stacked on top of The parameters of auto-encoder and auto decoder can be learnt for
layers. Each layer consists of multiple nodes which feed into the next recognition of cattle. The parameters are optimized by utilizing the
layer [30]. unsupervised training data. Then the output of the hidden layer is used
The basic working model of Restricted Boltzmann Machines (RBMs) as the feature for image representation. The auto-encoders is arranged
framework for cattle recognition is that a Deep Belief Network (DBN) to form an underground network by nursing the latent representation
deep learning framework is used for extracting and learning the ex- (output code) of the denoising auto-encoder found on the layer below
tracted set of texture features of muzzle point images. The DBN is the as input to the current layers [32].
graphical deep learning model which is used to learn the extracted The primary motivation to applied the stacked denoising auto-en-
features of muzzle images for hierarchical representation of the training coder technique are to reduce the noises and other artifacts from the
datasets by stacking of Restricted Boltzmann Machines (RBMs) classi- muzzle point image pattern The image of muzzle point pattern are
fication technique for classification and identification of individual captured from the unconstrained environment which are accompanied
cattle based on extracted muzzle point features. by variances in illumination, occlusion (covering and non-covering
during vegetation or body movement), body dynamics (due to head
5. Stacked denoising auto-encoder movement), and image blurriness, etc. Compared to the denoising auto-
encoder, these test muzzle point images are seen as clean data and these
Stacked Denoising Auto-encoder (SDAE) is encoding technique to test images can be seen as corrupted data. For robust recognition of
encode and decode the extracted features [31]. In this paper, Stacked cattle, we have applied the SDAE technique to learn the muzzle point
Denosing Auto-encoder (SDAE) deep learning technique is applied to features which are robust to these variances. The success of denoising
encode and decode the extracted texture feature extraction of muzzle auto-encoder convinces us of the possibility to learn such features.The

Fig. 7. Illustrates the architecture of DBN learning model composed of 4 stacked RBM classification approaches.

8
S. Kumar et al. Measurement 116 (2018) 1–17

unsupervised pre-training of stacked denoising auto-encoder archi- cattle recognition system. For supervised fine-tuning, a softmax (0,)
tecture is performed one layer at a time. Each layer is trained as a activation based layer with 35 nodes has been incorporated on top of
denoising auto-encoder by mitigating the error in reforming its input each stacked denoising auto-encoders. The four dropout [7] modules
(which is the output code of the previous layers). Once the first (K ) are applied in between layers to prevent over fitting. We have chosen
layers are trained, we can train the (K + 1) . The layer because we can the dropout probability as 0.35.
now calculate the code or latent representation from the layer below.
The non-linear mapping function (fW ,b ) is applied for vectorized input 6. Pre-Training and generalizability of proposed recognition
image (X ) and the hidden representation (Y ) is calculated as follows model
(shown in Eqs. (6), (7) and (8), respectively):
Y = fWij,b (X ) = S (Wij. X + b) In this subsection, pre-training of a proposed deep learning re-
(6)
cognition model is illustrated in detail. Deep learning architecture
Where S (·) denotes the sigmoid activation function and W and b are the needs the massive amount of datasets for the training the proposed
weight matrix and bias of the mapping function and Wij represents the models because the optimization functions are formulated to diminish
weight of the connection from the (ith) input node to the (jth ) hidden the training error of the proposed model. The proposed design tends to
node. The decoder maps the learnt features to the data space, using the learn maximum information from the given muzzle point image da-
following Eq. (6). A reconstruction step (GW ′,b′) is implemented on the taset.
lower dimensional mapping (y) as follows: The second major challenge is generalizability [28–30]. It provides
a method to train the model which can improve the generalizability
(GW ′,b′) = (Y ″) = S (W ′. Y + b′) (7)
without reducing the power of the proposed deep learning model. It is
In the Eq. (7), W ′ represents the weight matrix. Wij shows the weight accomplished by including a penalty variable to loss function, known as
function of the connection from the (ith) hidden node to the (jth ) decoder the regularizer function.
output node, and b′ presents the bias constraint of mapping function. The primary objective to apply the regularization is to provide the
Next, we have evaluated loss function of an auto-encoder as follows better representation of the information in the feature space to avoid
(show in Eq. (8)): over-fitting of a given problem, and yield to a solution faster by offering
ancillary features.
fae = ‖X −Y ″‖2F = ‖X −S (W ′S (WX + b) + b′)‖2F (8) The pre-training of each auto-encoder is implemented in greedy
The stacked denoising auto-encoder is a nonlinear auto-encoder for fashion one layer at a time. Each layer of an auto-encoder is trained by
representation of extracted feature which is different from PCA [33], lessening the rebuilding of its input. The performances of the above
and Linear Discrimination Analysis (LDA) [34] techniques. It has been three frameworks over a set of test images using the features obtained
decided that training an auto encoder is to reduce reconstruction error are calculated, individual, and the one which gives relatively better
by maximizing a lower bound on the mutual feature between input performance is employed for the recognition purpose. The various
layer representation and the learned representation [28,29]. We added identification settings and standard protocols are used for evaluating
the sparsity constraints for further boost the ability of auto-encoder for the experimental results of proposed system and existing handcrafted
representation of extracted muzzle point images in proposed deep texture feature descriptor techniques. The experimental results and
networks. The proposed deep learning approach enhances the gen- discussion of proposed system is illustrated in the next section.
eralization of auto-encoder by training with locally corrupted inputs of
muzzle point image pattern. 7. Experimental results and discussions
In denoising auto-encoder, input (X ) is first corrupted by some pre-
∼
defined noise, for example, Additive Gaussian noise (X |X ∼ N (x ,σ 2I )) , In this section, we have performed the experiments to compute the
masking noise (a fraction of X is forced to 0), or salt-and-pepper noise (a effectiveness of the proposed deep learning approach for the recogni-
fraction of X is forced to be 0 or 1). Fig. 8(a) and (b) represents the 20% tion of cattle using muzzle point image pattern. The comparison with
and 30% corrupted muzzle point images for better representation of existing benchmark algorithms (texture feature descriptor technique,
extracted muzzle point images. We have applied the Rectified Linear appearance based feature extraction and representation, and learnt
Unit (ReLU) activation function in all the encoder layers of proposed feature techniques) is accomplished to evaluate the identification

Fig. 8. (a) Illustrates 20% corrupted images of muzzle point pattern (b) shows the 30% corrupted muzzle point images using stacked denoising auto-encoder.

9
S. Kumar et al. Measurement 116 (2018) 1–17

accuracy in multiple identification settings. point image is shown in Figs. 9 and 10, 11, 12, and 13 respectively.
For performance evaluation of experimental results, the database of Moreover, the appearance based feature extraction, and re-
muzzle point image pattern is segmented into following phases: (1) presentation methods, and learned feature based techniques are used
training phase and (2) testing phase. In the training phase, 100 muzzle for the evaluation and comparison of the experimental results of pro-
point images (10 cattle (subject) ×10 image of each subject) are uti- posed approach.
lized to train the proposed deep learning approach. In the testing phase, For comparative analysis, we have also applied the appearance
400 testing pairs (40 cattle (subject ×10 image of each subject)) of based face recognition and representation approaches, such Principal
muzzle point image pattern in each fold are used to test the probe Component Analysis (PCA) (Eigen-values) [33], Linear Discriminant
images. Analysis (LDA) [40], Kernel-LDA [41], and Direct-LDA [42] for the
The deep learning based framework demands the massive amount of evaluation of experimental results. Dense-SIFT (DSIFT) feature de-
database to train the proposed network. Despite, the size of muzzle scriptor technique is applied to extract the set of key points of muzzle
point image database is 5000 images which are a relatively smaller point images. These key points are distributed at regular intervals on a
image database. It is not satisfactory to adequately train a deep belief uniform grid. In each grid (cell), the discriminatory key-points are se-
network or a stacked denoising auto-encoder. Therefore, a transfer lected from a descriptor of length n × 16, where n is the number of
learning approach is applied for the fine tuning the weight between the orientations.
input and hidden layer and determined the pre-training the proposed For the evaluation of performance, we have performed the three
deep learning approach. experiments to evaluate the effectiveness of the proposed approach and
compare with existing benchmark texture descriptor technique and
appearance based feature extraction and representation algorithms in
7.1. Performance evaluations multiple identification settings. In this section, we have illustrated the
detail evaluation of the experimental results of cattle recognition as
The proposed deep learning based approach for cattle recognition is follows:
motivated by the observation that the muzzle point image pattern of
cattle consists of rich and dense texture features in the form of bead and 7.1.1. Experiment 1: Accuracy analysis of different state-of-the-art deep
ridges pattern. The bead and ridge pattern of muzzle point image are learning approaches
salient biometric features for cattle recognition. Muzzle point images of 100 cattle (subject) are randomly chosen for
Moreover, it is challenging to restrict the body movement due to training the system and the remaining muzzle point images corre-
head movement and unconstrained environments during the data ac- sponding to 400 (cattle) are used for testing with 1, 2, 3, and 4 images
quisition of cattle. per subject in the gallery. The least one shot similarity based matching
During data acquisition, the unconstrained environments such as scores are obtained per subject. It is used as the similarity match score.
low illumination, poor image quality, and blurriness due to head Further, all the experimental results are reported with five-times
movement are also the biggest problems for recognition of individual random sub sampling based cross validation technique. The identifi-
cattle, implying that appearance based feature extraction and re- cation experiments are performed and the results are reported in terms
presentation algorithms may not yield good results. of rank-1 identification accuracy along with Cumulative Match
On the other hand, local feature based feature extraction and re- Characteristics (CMC) curves as follows:
presentation algorithm can provide good results. The hypothesis is that
feature (information) content present in the muzzle point image pattern
varies with unconstrained environments for cattle.
• The experimental results are also analyzed with varying gallery sizes
and summarized in Tables 1 and 2. The experimental results are also
To efficiently extract and encode texture features of muzzle point depicted in Figs. 14 and 15, respectively. With the proposed deep
image pattern of cattle, local feature descriptor techniques must be learning approaches, such as CNN, SDA and DBN yield 75.98%,
used. In this paper, the bead and ridges texture features are extracted 88.46%, and 95.99% identification accuracy for identification of
from the muzzle point image pattern of cattle database using the individual cattle, respectively.
handcrafted texture feature descriptor techniques. The handcrafted
texture descriptor features and learnt features techniques used for
• The proposed approach has two components: (1) learning the robust
texture feature of muzzle point image feature for representation and
feature extraction and representation. These approaches are mainly learning the distance learning based metric with One-Shot Similarity
Local Binary Pattern (LBP) [35], Circular-LBP [36], Scale Invariant (OSS) based similarity matching techniques.
Feature Transform (SIFT) [37], Dense-SIFT [37], Speeded Up Robust
Feature (SURF) [38], and Vector of Locally Aggregated Descriptors
• For evaluation of effectiveness of both the components of proposed
approach, we tactically replaced one component at a time with ex-
(VLAD) techniques [39]. The computation of features and encoding of isting descriptors or matchers (Chi-square (χ 2 ) based dissimilarity
the local binary pattern based feature, SIFT and SURF feature of muzzle

Fig. 9. Illustrates the SIFT keypoint descriptor: (a) input muzzle point image (b) A SIFT descriptor of the size (m × n) is chosen from muzzle point image, (c) selection of 16 × 16 pixel
orientations, (d) 4 × 4 cells descriptor with 8 pixel orientations are chosen. The size of single SIFT keypoint descriptor is 4 × 4 × 8 = 128 element.

10
S. Kumar et al. Measurement 116 (2018) 1–17

Fig. 10. Illustrates (a) the process of SIFT keypoint localization; detection and (b) matching of test muzzle point image with stored muzzle point image using SIFT keypoint descriptor.

Fig. 11. Illustrates the extraction and encoding of local binary pattern based descriptor features from the muzzle point image pattern.

Fig. 12. Illustrates the process of building of SURF descriptor: (a)-(b) shows the detection of keypoints in the muzzle point images (c) matching of muzzle point images based on keypoint
SURF descriptor with size (for a neighborhood of size 6s where s is scaling parameter of wavelet responses in horizontal and vertical directions.

matching technique) and compared the experimental results with framework are calculated by considering different number of muzzle
four gallery muzzle point images per subject (cattle). point image features amongst the calculated ones. It can be observed
• As shown in the Table 1, it can be observed that with increasing the that with the increase in usage of number of texture features of
number of muzzle point texture feature of cattle database per sub- muzzle point image pattern of cattle, the identification accuracies
ject from one to four, the identification performance of deep are gradually increasing in all the three deep learning algorithms (as
learning framework using the DBN learning technique provides the shown in Table 2). This explains the importance of all the muzzle
highest identification accuracy for identification of individual cattle. point image features.
Table 1 illustrates the average identification accuracy based on in-
dividual patches of the muzzle point image pattern of cattle for re- As explained earlier, as the number of features is increasing, the
cognizing individual cattle. identification accuracies are gradually increasing but at every instant
• Identification accuracies using the proposed deep learning DBN technique provided the good experimental results amongst the

11
S. Kumar et al. Measurement 116 (2018) 1–17

Fig. 13. Matching of keypoints of muzzle point images using SURF descriptor.

Table 1
Illustrates the identification accuracy (%) of CNN, SDAE, and DBN deep learning ap-
proaches.

S. No. Proposed approach Identification accuracy (%)

1 CNN 75.98
2 SDAE 88.76
3 DBN 95.99

Table 2
Identification accuracy (%) of CNN, SDAE and DBN deep learning approaches.

Number of feature sets Identification accuracy (%)

CNN SDAE DBN

50 63.75 67.75 65.95

100 67.98 68.65 69.85
150 73.85 71.96 75.85
200 76.75 76.92 77.94 Fig. 15. CMC curve illustrate the identification accuracy vs. number of patches (images).
250 79.98 78.67 82.99
300 82.99 85.98 86.92
350 86.75 89.79 94.75 learning based DBN approach yields 98.99% of identification accuracy
400 95.98 96.92 98.99 for identification of cattle.
It is observed that the identification accuracies of the three deep
learning frameworks are gradually increasing with the increase in the
number features of selected patches from the muzzle point texture
features. After the selection of 400 numbers of features from each patch
(size: 200 × 200 pixels), deep learning algorithms provide the robust
representation of muzzle point features in the different layer of the
proposed framework.
When the size of each patches of muzzle point images are reduced,
the selected discriminatory set of the muzzle texture feature has also
reduced, therefore, we have selected the 400 number of features as
better muzzle point feature sets from the extracted patches. This shows
us that all the patches of muzzle point image pattern have been cal-
culated collectively to describe and represent the discriminatory set of
muzzle point features to the best extent than a set of few patches (as
shown in Table 1 and Table 2, respectively).

7.1.2. Experiment 2:Average accuracy analysis with standard deviation of

existing handcrafted texture feature and representations
Fig. 14. CMC curve illustrate the identification accuracy vs. number of muzzle point
features. In this section, we have evaluated the performance of the existing
handcrafted texture feature and representation algorithms.

three deep learning approaches. This is can be seen in Fig. 14.

The identification accuracy of DBN is higher as compared to con- • We have also applied the handcrafted texture features based re-
presentation algorithms for evaluations of experimental results for
volution neural network and stacked denoising auto-encoder based
identification of cattle. The experimental results are summarized in
deep learning approaches for recognition of individual cattle. The deep

12
S. Kumar et al. Measurement 116 (2018) 1–17

Table 3 discriminatory keypoints of muzzle point image pattern efficiently

The experimental results are reported in terms of average accuracy with standard de- in the different scales (vertical and horizontal scale of wavelet re-
viation over 10-fold cross-validation for Existing Handcrafted Texture Features based
sponses) as compared to SIFT and Dense-SIFT approaches.
methods.

Existing Handcrafted Methods Number of muzzle point images per subject 7.1.3. Experiment 3: Average accuracy analysis with standard deviation of
Texture Features (cattle) in gallery appearance based face recognition algorithms
based method 1 2 3 4
72.84 73.52 73.76 74.97
Features of muzzle point images were extracted using the appear-
LBP+ χ 2
(0.8) (0.83) (1.23) (1.2) ance based feature extraction and representation approaches, such as
Circular- 73.87 76.5 78.96 79.87 PCA, 2-D PCA, LDA, and its modified LDA version (e.g., Kernel-LDA,
LBP (1.5) (1.83) (1.56) (2.2) Direct-LDA) algorithms, respectively. As shown in Fig. 2, following
SIFT+ χ 2 67.98 69.97 70.97 72.85
operations are applied to perform matching of extracted set of features
(1.2) (1.2) (2.2) (1.3)
Dense- 69.75 71.84 73.98 75.89 and representation of muzzle point texture feature for identification of
SIFT+ χ 2 (2.6) (2.6) (1.4) (1.65) individual cattle:
SURF 75.48 78.99 85.49 89.76
(0.9) (1.5) (1.3) (1.43)
• Principal Components Analysis (PCA) technique is used to perform
the dimensionality reduction of extracted feature of muzzle point
image database. PCA technique compute the on the feature space.
Table 3, respectively.
Principal components corresponding to 99% Eigen-values in the
• In Table 3, the existing handcrafted texture feature based descriptor PCA subspace are retained.
algorithms such as Local Binary Pattern (LBP) and Circular-LBP
(Circular-LBP) feature descriptors provide the rank-1 identification • The extracted texture features are classified is using supervised
learning Linear Discriminant Analysis (LDA) technique.
accuracy of 72.84 ± 0.80% to 74.97 ± 1.2 and 73.87 ± 1.5 to
79.87 ± 2.2, respectively with 4 muzzle point images of cattle as • The Cosine similarity matching based method, One-Shot Similarity
(OSS), distance learning based metric technique are used for the
gallery image per subject (cattle).
similarity matching score after matching a pair of samples (muzzle
• The SIFT feature descriptor techniques such as SIFT and Desne-SIFT
point image patterns of cattle).
(D-SIFT) yield identification accuracy of 69.75 ± 2.6 to72.85 ±
1.3 and 69.75 ± 2.6 to 75.89 ± 1.65, respectively. • After that the similarity matching scores are measured. The identi-
fication accuracies of cattle are shown in Table 5. Based on ob-
• The evaluation of the experimental results, we performed the testing
servation of Table 5 it is shown that the Kernel-LDA and Direct-LDA
of muzzle point images multiple in the gallery. During the experi-
techniques provide 60.89 ± 1.87% to 68.97 ± 1.28% and
mental results on every test muzzle point images, we have calcu-
63.77 ± 1.79 to 69.97 ± 1.33, identification accuracies for cattle
lated the least distance matching scores for every subject.
recognition.
• The matching score values are used for evaluation of the experi-
• On the other hand, texture feature descriptor techniques, such as
mental results. The experimental results illustrate that even with
SURF, LBP, Circular-LBP, SIFT, Dense-SIFT and VLAD are hand-
multiple gallery muzzle point images, the identification accuracies
crafted existing features based descriptors for the better re-
are reported in the increasing order however, the performance
presentation of muzzle point features for identification of individual
leaning remains the same.
cattle. The LBP and Dense-SIFT descriptor algorithms illustrate the
• The better performances of texture feature based descriptors algo-
low recognition accuracy as compared to VLAD + LDA with One-
rithms are attributed to spatial collation in regional blocks that is
Shot-Similarity (OSS) and VLAD + LDA with Support Vector
able to good deal with the covariates, such as pose due to head
Machine (SVM) techniques for identification of individual cattle
movement, body dynamics, poor image quality and low illumina-
using their primary muzzle point image pattern. The learnt feature
tion.
descriptor techniques, such as VLAD + LDA + OSS and
• Vector of Locally Aggregated Descriptors (VLAD) [39] is a feature
VLAD + LDA + SVM technique yield 60.89 ± 1.5% to
extractor learning based descriptor algorithm. It extracts visual in-
59.64 ± 1.12% and 50.76 ± 1.6 to 67.98 ± 1.17 identification
formation (features) the training datasets [39].
accuracy, respectively.
• As illustrated in given Table 4, the better recognition accuracy is
achieved with OSS [43] (SVM [44] classification model) for this
The capture muzzle point image database has several significant
representation when the gallery consists of four images per cattle
challenges due to the unconstrained environment, such as poor illu-
subject, with rank-1 accuracy of 45.98 ± 0.8% to 59.64 ± 1.12%
mination, image quality, low contrast, and blurred. Therefore, existing
and VLAD + LDA + SVM technique yields 50.76 ± 1.6% to
feature extraction and representation algorithms are unable to perform
67.98 ± 1.17%, respectively.
identification of cattle using muzzle point images. The learning-based
• Based on overall performance of the existing handcrafted feature
descriptors techniques, SURF descriptor provides better identifica-
Table 5
tion accuracy of 75.48 ± 0.9) to 89.76 ± 1.43, respectively be- The experimental results are reported in terms of average accuracy with standard de-
cause the detection of the SURF descriptor compute the viation over 10-fold cross-validation on holistic features based recognition algorithms.

Holistic Feature Methods Number of muzzle point images per subject (cattle)
Table 4 based in gallery
The experimental results are reported in terms of average accuracy with standard de- Methods 1 2 3 4
viation over 10-fold cross-validation for texture holistic features based recognition al- PCA 65.89 68.95 69.86 70.97
gorithms. (1.67) (1.83) (1.23) (1.2)
2D-PCA 67.89 69.65 73.86 75.67
Existing Learnt Methods Number of muzzle point images per subject (1.84) (1.63) (1.51) (1.45)
Feature (cattle) in gallery LDA 69.96 72.91 73.48 74.99
based 1 2 3 4 (1.98) (1.64) (1.43) (1.37)
method VLAD 45.98 49.89 53.94 59.64 Kernel-LDA 60.89 65.65 67.86 68.97
+LDA+(OSS) (1.5) (1.47) (1.22) (1.12) (1.87) (1.73) (1.57) (1.28)
VLAD 50.76 54.92 58.74 (1.17) Direct- LDA 63.77 65.96 68.96 69.97
+LDA+SVM (1.6) (1.27) (1.02) (1.79) (1.63) (1.53) (1.33)

13
S. Kumar et al. Measurement 116 (2018) 1–17

Table 6
Comparison of our proposed approach with the literature.

Authors No. of images (Muzzle print Technique used Identification accuracy (%)
images) and cattle

Noviyanto et al. [4] 80 images SURF+kappa statistic+Eigen-face Algorithm 89.30%

Minagawa et al. 43 muzzle print images using ink PCA + Eigen-values based approaches 30%
[18]
Barry et al. [19] 29 cattle breeds Eigen-value+ Segmentation based technique 98.50%
Awad et al. [20,50] 15 cattle breeds SIFT+RANSAC 93.30%
Noviyanto et al. 48 muzzle images SIFT+PCA 0.0167 (EER)
[21]
Kumar et al. [4,23] 300 cattle PCA+ LDA+ICA 85.95%
Gaber et al. [24] 31 cattle WLD+ABD 99%
Cai and Li [45] 30 cattle RASL + WLBP 95.30%
Tharwat et al. [46] 31 cattle Gabor + SVM 99.50%
Kumar et al. Pet animals (dog), PCA+LBP +modified algorithms (Batch-ILDA, CCIPCA, Incremental-SVM) 94.86%
[47,54] 50 dog breeds
This research study 5000 muzzle point image (500 Deep learning approaches (Convolution Neural Network (CNN) + Deep Belief 95.98% (CNN), 95.99%(DBN)+
subject and each subject has 10 Network (DBN) and Stacked Denoisingauto-encoders (SDAE))+SVM+One-Shot- 96.92%(SDAE),98.99%
images) Similarity (OSS) (DBN)

Where, WLD = Weber’s Local Descriptor, ADB = AdaBoost classifier, WLBP = Weber’s Local Binary Pattern Descriptor, RANSAC = RANdom Sample Consensus algorithm.

feature extraction and matching approaches cater an explicit encoding In the classical muzzle print identification based systems includes
mechanism of extracted feature to improve the recognition accuracy of well-defined manual procedures for identification of cattle. These sys-
individual cattle. tems have provided the paradigm to the interdisciplinary researches,
We have applied the One-Shot Similarity (OSS) similarity using scientists, and veterinary professionals for identification and mon-
Fisher linear discriminant analysis (FLDA), and incremental support itoring of individual cattle. However, muzzle print-image-based iden-
vector machine (SVM) classification models are applied. The 1-class tification system consumes more time for processing of images, feature
online incremental SVM (1-online ISVM) model is used to classify the extraction and image analysis works. Therefore, in this research work,
extracted texture features of muzzle point images. we have captured the image of muzzle point pattern of cattle using 30-
The OSS matching is a semi-supervised based matching similarity megapixel digital camera.
technique. It selects the unlabeled training data as a set of negative To perform the identification of cattle, the muzzle point image da-
constraints against which two input sample images of muzzle point tabase is prepared from the Department of Dairying and Husbandry,
pattern are matched. Based on the overall experimentations, and Institute of Agriculture Sciences (I.A.S), Banaras Hindu University
achieved recognition accuracy our belief is that handcrafted existing (B.H.U), Varanasi, India-221005.
benchmark based texture features extraction and representation have The prepared database consists of original images of muzzle point
are bound and limited representation capacity as these descriptors are pattern of cattle. The database also includes various covariates of
not applicable for better recognition of cattle in the specific problem muzzle point images due to poor image quality, low illumination, pose,
domain. variation images.
The proposed deep learning based framework performs the en- In the available literature, there is no availability of muzzle point
coding and learning of the discriminant feature representation of image database in the public domain. Very few researches have been
muzzle point image pattern and the machine learning based distance done for the identification of cattle based on the muzzle point image
metric techniques capture the semantic representation of features and pattern. Based on printed muzzle images, we have compared the ex-
understanding of the SDAE encoding and decoding scheme. perimental results of proposed approach with previously published re-
sults for identification of cattle based on muzzle print image database.
7.2. Comparative analysis The experimental results are achieved by applying the computer vision,
image processing and pattern recognition techniques. The comparative
In this section, we have performed the detailed analysis of experi- analyses of the experimental results are shown in Table 6.
mental analysis of proposed deep learning approach for recognition of Noviyanto et al. [4] proposed a method using speeded up robust
cattle based on muzzle point images. We now compare the performance features and Eigen-face based approaches for recognition of individual
of our proposed deep learning framework based system against current cattle based on muzzle print image database. The major shortcoming of
state-of-art approaches for recognition of livestock based on their proposed approach is that the authors do not include the image filtering
muzzle print images from the literature. techniques to remove the noise from the captured muzzle print images.
For, capturing the muzzle image pattern of cattle, the manual ac- It may affect the identification accuracy of proposed system. Further-
quisition methodologies are not applied. In the classical acquisition more, they performed the experimentations on the small database of
methodology of muzzle print images of cattle includes the equipment muzzle print image of cattle. The proposed approach cannot test on the
and materials (such as A5-size white papers, black ink and stamp (im- different rotation and scaling of muzzle print images for identification
pacted), soft cottons, hard ropes, tissue paper and assistant team of the of cattle.
dairy staff members) to capture the images of muzzle print of individual In [18], authors proposed a cattle recognition based framework for
cattle for preparation of database. identification of cattle using Eigen-values based approaches. The au-
The print images of muzzle pattern are captured on A-5 white paper thors have applied principal component analysis techniques to mitigate
with blue ink. The captured print images are required to mitigate the the dimensionality of extracted features.
noises and other artifacts using image processing techniques. However, The proposed approach by Minagawa et al. [18] has not reported
the captured print images have very low image quality. Therefore, it exactly experimental results the same due to the unexplained filtering
also requires some enhancement and transformation techniques to techniques. In [19], Barry et al. proposed cattle identification using
convert the print muzzle image into digital muzzle print images (300 principal component analysis and Euclidean distance classifier techni-
dot per inch (DPI), resolution of images). ques based on muzzle print images.

14
S. Kumar et al. Measurement 116 (2018) 1–17

The proposed approach is selected for training separately on a dif- identifiers. The proposed recognition system apply the one-shot simi-
ferent number of normalized muzzle images sets of 2, 4, 6, 8, and 10 larity matching techniques to calculate the similarity matching score
training images from 29 cattle. after the matching of query face image of cattle with stored image
The drawback of the approach is that an experimental result was datasets of animal. Moreover, the distance metric based learning
taken on a separate set of 3 images only per animal. The authors have method is used for matching and classifying the extracted features of
not performed any cross-validation of the experimental results. The face images for recognition of pet animals (dog). The efficacy of pro-
implementation of the approach proposed by Barry et al. [19] is also posed recognition system yields 96.87% recognition rate.
not exactly the same due to the watershed segmentation technique is Based on overall observation, we conclude that our proposed
not implemented properly. The approach proposed by Barry et al. [19] method uses the deep learning based framework for recognition of
has been very strict due to the false match has been zero. 468 false non- cattle and achieves significant improvements over all of this existing
matches have been reported over 560 genuine matching. handcrafted feature descriptor, baselines algorithms which are avail-
In the similar direction, Awad et al. [20] proposed a method to able in the literature [55–62].
improve the performance of proposed cattle identification system. They
have applied the SIFT keypoint matching based descriptors. The SIFT 8. Conclusions and future directions
descriptor technique is utilized to find out the keypoints for matching of
muzzle print images. For better identification, Random Sample Con- In this paper, we proposed a novel deep learning approach to
sensus (RANSAC) technique is utilized with the SIFT to mitigate the identifying the individual cattle using muzzle point image pattern. The
noises such as outlier points. However, the proposed approach has deep learning approach is applied to learn a discriminatory feature
following major limitations-(1) no cross-validation of experimental re- representation of muzzle images with limited training dataset. With the
sults, (2) identification accuracy is suffered from the noises such out- proposed deep learning approaches, such as CNN, SDA and DBN yield
liers, blurriness, and poor image quality. Noviyanto et al. [21] im- 75.98%, 88.46%, and 95.99% identification accuracy, respectively.
plemented the proposed a matching refinement technique based on The handcrafted texture features based representation algorithms
SIFT keypoints matching technique. are utilized for evaluations of experimental results. The Local Binary
In [24], authors proposed a method for identification of cattle using Pattern (LBP) and Circular-LBP (Circular-LBP) feature descriptor based
Weber’s Local Descriptor (WLD) technique. The proposed approach technique provide the rank-1 identification accuracy of
extract the features from cattle muzzle print images (images from 31 16.80 ± 0.80% to and 26.97 ± 1.2%, respectively with four muzzle
head of livestock). The extracted features are classified by AdaBoost point images as gallery image per subject (cattle). In the case of ap-
classification model to identify the head of individual cattle from their pearance based feature extraction and representation approaches, such
WLD descriptor features. The limitation of this paper is that experi- as principal components analysis is used to perform dimensionality
mentations are carried out the small datasets, and no cross-validation reduction on the feature space.
technique is applied to validate the experimental results. The identification of cattle based on their muzzle point images is
Cai and Li [45] proposed a method for automatic recognition of performed using Linear Discriminant Analysis (LDA) with One-Shot
cattle using local binary pattern based facial feature descriptor and Similarity (OSS) technique. It is used to match a pair of samples and
extended LBP descriptors techniques. The robust alignment by sparse generate the match scores. The identification accuracies are shown in
and low-rank decomposition approaches is applied to align the face Table 5, respectively. It can be observed that Direct-Kernel-LDA pro-
images of cattle due to poor illumination, image misalignment and vides the 15.89 ± 1.7% to 29.97 ± 1.13% identification accuracy.
occlusion problem in the test face image of cattle. The learnt feature descriptor techniques, such as VLAD+LDA+OSS
The dissimilarity between test and trained images are performed on and VLAD+LDA+SVM techniques yield 45.98 ± 1.5% to
a separate set of face images using the weighted Chi-square distance 59.64 ± 1.12% and 50.76 ± 1.6 to 67.98 ± 1.17 identification ac-
technique. The major shortcomings of this paper are that authors have curacy, respectively. Based on observation, we conclude that deep be-
not performed the experimental results on slight datasets, (2) no pre- lief network deep learning approach provides better identification ac-
processing technique is applied for processing of facial images of cattle curacy for recognition of individual cattle. Hence it can be concluded
using image processing technique and (3) any cross-validation is not that the DBN based framework is the right choice for recognition pur-
applied to verify the identification accuracy. In [46], the authors pro- pose.
posed a cattle identification using Gabor filter-based feature extraction For further improvisation of deep learning based recognition fra-
technique. The proposed extract Gabor features from muzzle print mework, the proposed framework can be implemented on the android
images. The extracted features are classified by using support vector platform that can easily available for smart or android devices for
machine based classification technique at different kernels. verification and identification and verification of false insurance claims
The author, Kumar et al. [48] proposed a real cattle recognition in real time scenario.
system using Fisher locality preserving projection based recognition We postulate that the traditional animal recognition methodologies
algorithm for the recognition of individual cattle in real time. In the and automatic animal recognition algorithms that are tailored specifi-
proposed systems captures the images of cattle using a surveillance cally for identification of cattle, via unambiguous training. It can be
camera and captured image of muzzle point of cattle is transferred them able to perform the recognition of cattle more efficiently. The proposed
to the server side by using wireless network communication technology deep learning based recognition of cattle caters a friendly, non-invasive,
[49]. robust as well as cost-effective solution using smart devices or low-cost
The proposed recognition algorithm based on muzzle point features camera for the identification of species or individual animals.
approach yields 96.87% accuracy for identification of cattle. The author In future, we plan to extend the proposed cattle recognition system
Mishra et al. [53] did the dermatoglypics of cattle muzzle pattern using for the identification of different animals in the real time. We would
blue ink images to identify bovines. In his study, they have performed like to include the following points as part of our future work:
the identification of cattle using classification techniques.
Recently, the authors Kumar et al. [51,52] proposed a real time • We would like to design multi-modal cattle recognition system using
monitoring and tracking systems for the pet animals in smart cities muzzle point image and face image of cattle for accurate identifi-
using animal biometrics and computer vision techniques. cation and verification in real-time.
The proposed system extracts the facial images based pixel intensity • We would like to increase the performance of the proposed cattle
feature for identification of pet animal (dogs). The systems uniquely recognition system using multi-modal system and feature fusion
identify the pet animal based on their primary animal biometric techniques. The fusion technique can be fused the discriminatory set

15
S. Kumar et al. Measurement 116 (2018) 1–17

of texture features of the muzzle point images with facial images of [26] S. Lawrence, C.L. Giles, A.C. Tsoi, A.D. Back, Face recognition: a convolutional
neural-network approach, IEEE Trans. Neural Netw. 8 (1) (1997) 98–113.
individual cattle. [27] A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep con-
• We postulate that it can be more helpful even depth level analysis of volutional neural networks, Adv. Neural Inform. Proc. Syst. (2012) 1097–1105.
[28] P. Vincent, H. Larochelle, Y. Bengio, P.A. Manzagol, Extracting and composing
experimental results of proposed multi-model based cattle recogni-
robust features with denoisingauto-encoders, in: Proceedings of the 25th ACM in-
tion system. ternational conference on Machine learning, 2008, pp. 1096–1103.
• Finally, we would like to increase the size of cattle database for [29] Y. Bengio, Learning deep architectures for AI, Found. Trends Mach. Learn. 2 (1)
(2009) 1–127.
validation of experimental results with benchmark existing hand-
[30] Y. Bengio, A. Courville, P. Vincent, Representation learning: A review and new
crafted texture descriptor techniques and deep learning based fea- perspectives, IEEE Trans. Pattern Anal. Mach. Intell. 35 (8) (2013) 1798–1828.
ture learning and representation techniques in the computer vision. [31] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, P.-A. Manzagol, Stacked denoi-
singauto-encoders: learning useful representations in a deep network with a local
denoising criterion, JMLR 11 (2010) 3371–3408.
References
[32] P. Vincent, H. Larochelle, Y. Bengio, P.A. Manzagol, Extracting and composing
robust features with denoisingauto-encoders, Proc. 25th Int. Conf. Mach. Learn.
[1] H.S. Kohl, T. Burkhart, Animal biometrics: quantifying and detecting phenotypic (2008) 1096–1103.
appearance, Trends Ecol. Evol. 28 (7) (2013) 432–441. [33] M.A. Turk, A.P. Pentland, Face recognition using eigenfaces, in: Proceedings of IEEE
[2] J. Duck, C. Finn, A. Hutcheon, P. Vera, J. Salas, S. Ravela, Sloop: a pattern retrieval Computer Society Conference on Computer Vision and Pattern Recognition (CVPR
engine for individual animal identification, Pattern Recogn. 48 (4) (2015) ’91), 1991, pp. 586–591.
1059–1073. [34] K. Etemad, R. Chellappa, Discriminant analysis for recognition of human face
[3] S. Kumar, S. Tiwari, S.K. Singh, Face recognition of cattle: Can it be done? Proc. images, JOSA A 14 (8) (1997) 1724–1733.
Natl. Acad. Sci., India, Sect. A 86 (2) (2016) 137–148. [35] T. Ahonen, A. Hadid, Pietikainen M Face description with local binary patterns:
[4] A. Noviyanto, A.M. Arymurthy, Automatic cattle identification based on muzzle Application to face recognition, IEEE Trans. Pattern Anal. Mach. Intell. 28 (12)
photo using speed-up robust features approach, in: Proceedings of the 3rd European (2006) 2037–2041.
conference of computer science, ECCS, 2012, vol. 110, p. 114. [36] T. Ojala, M. Pietikainen, Multiresolution gray-scale and rotation invariant texture
[5] S. Kumar, S.K. Singh, T. Dutta, H.P. Gupta, Poster: a real-time cattle recognition classification with local binary patterns, IEEE Trans. Pattern Anal. Mach. Intell. 24
system using wireless multimedia networks, in: Proceedings of the 14th Annual (7) (2002) 971–987.
International Conference on Mobile Systems, Applications, and Services [37] D.G. Lowe, Object recognition from local scale-invariant features, in: Proceedings of
Companion, Singapore, June 2016, pp. 48–48. the 7th IEEE international conference on Computer vision, 2009, 2, pp. 1150–1157.
[6] D. Wardrope, Problems with the use of ear tags in cattle, Veterinary Record (United [38] H. Bay, A. Ess, T. Tuytelaars, L. Van Gool, Speeded-up robust features (surf), Comp.
Kingdom) 37 (26) (2009) 675. Vision Image Understand. 110 (3) (2008) 346–359.
[7] S. Kumar, S.K. Singh, Visual animal biometrics: survey, IET Biometrics. 6 (3) (2016 [39] M. Jégou, H. Douze, C. Schmid, P. Pérez, Aggregating local descriptors into a
Nov 30) 139–156. compact image representation, Proc. IEEE Conf. Comput. Vis. Pattern Recognit.
[8] S. Kumar,S.K. Singh, T. Datta, H.P. Gupta, A fast cattle recognition system using (2010) 3304–3311.
smart devices, in: Proceedings of the 2016 ACM conference on Multimedia, [40] P.N. Belhumeur, J.P. Hespanha, D.J. Kriegman, Eigenfa.ces vs. fisherfaces:
Amsterdam, The Netherlands, October 15-19, 2016, pp.742–743. Recognition using class specific linear projection, IEEE Trans. Pattern Anal. Mach.
[9] W. Petersen, The identification of the bovine by means of nose-prints, J. Dairy Sci. 5 Intell. 19 (7) (1997) 711–720.
(3) (1922) 249–258. [41] G. Baudat, F. Anouar, Generalized discriminant analysis using a kernel approach,
[10] A.M. Johnston, D.S. Edwards, Welfare implications of identification of cattle by ear Neural Comput. 12 (10) (2000) 2385–2424.
tags, Vet. Rec. 138 (25) (1996) 612–614. [42] K.R. Muller, S. Mika, G. Ratsch, K. Tsuda, B. Scholkopf, An introduction to kernel-
[11] U.G. Barron, F. Butler, K. McDonnell, S. Ward, The end of the identity crisis? based learning algorithms, IEEE Trans Neural Netw. 12 (2) (2001) 181–201.
Advances in biometric markers for animal identification, Irish Veterinary J. 62 (3) [43] L. Wolf, T. Hassner, Y. Taigman, The one-shot similarity kernel, Proc. IEEE 12th Int.
(2009) 204–208. Conf. Comput. Vis., Sep. /Oct. (2009) 897–902.
[12] Z. Wang, Z. Fu, W. Chen, J. Hu, A rfid-based traceability system for cattle breeding [44] B. Schölkopf, R.C. Williamson, A.J. Smola, J. Shawe-Taylor, J.C. Platt, Support
in china, in: Proceedings of 2010 IEEE International Conference on Computer vector method for novelty detection, Proc. Neural Inf. Process. Syst. 12 (1999)
Application and System Modeling (ICCASM), 2010, 2, pp. V2–567. 582–588.
[13] A. Krizhevsky, I. Sutskever, G. Hinton, ImageNet classification with deep con- [45] C. Cai, J. Li, Cattle face recognition using local binary pattern descriptor, in:
volutional neural networks, in Advances in Neural Information Processing Systems, Proceedings of IEEE Asia-Pacific Signal and Information Processing Association
Curran & Associates Inc., Red Hook, NY, USA, 2012, pp. 1097–1105. Annual Summit and Conference (APSIPA), pp. 1–4.
[14] C. Farabet, C. Couprie, L. Najman, Y. LeCun, Learning hierarchical features for [46] A. Tharwat, T. Gaber, A.E. Hassanien, Cattle identification based on muzzle images
scene labeling, IEEE Transaction on Pattern Analysis Machine Intelligence 35 (8) using gabor features and SVM classifier, Int. Conf. Adv. Mach. Learn. Technol. Appl.
(2013) 1915–1929. (2014) 236–247.
[15] Y. Sun, X. Wang, X. Tang, Deep convolutional network cascade for facial point [47] S. Kumar, S. Singh, Biometric recognition for pet animal, J. Softw. Eng. Appl. 7
detection, Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (2013) 3476–3483. (2014) 470–482, http://dx.doi.org/10.4236/jsea.2014.75044.
[16] A. Baranov, R. Graml, F. Pirchner, D. Schmid, Breed differences and intra-breed [48] S. Kumar, S.K. Singh, Automatic identification of cattle using muzzle point pattern:
genetic variability of dermatoglyphic pattern of cattle, J. Anim. Breed. Genet. 110 a hybrid feature extraction and classification paradigm (2016), http://dx.doi.org/
(1–6) (1993) 385–439. 10.1007/s11042-016-4181-9.
[17] S. Mishra, O. Tomer, E. Kalm, Muzzle dermatoglyphics: A new method to identify [49] S. Kumar, S.K. Singh, R.S. Singh, A.K. Singh, S. Tiwari, Real-time recognition of
bovines, Asian Livestock, FAO Publication, 1995 vol. XX, No. 8. cattle using animal biometrics, J. Real-Time Image Proc. 1–22 (2016).
[18] H. Minagawa, T. Fujimura, M. Ichiyanagi, K. Tanaka, and Mei Fangquan, [50] A.I. Awad, From classical methods to animal biometrics: A review on cattle iden-
Identification of beef cattle by analyzing images of their muzzle patterns lifted on tification and tracking, Comp. Electr. Agr. 123 (2016) 423–435.
paper, in: Proceedings of the 3rd Asian Conference for Information Technology in [51] S. Kumar, S.K. Singh, Monitoring of pet animal in smart cities using animal bio-
Asian agricultural information technology & management, 2002, pp. 596–600. metrics, Future Gener. Comp. Syst. (2016 Dec 14).
[19] B. Barry, U. Gonzales-Barron, K. McDonnell, F. Butler, S. Ward, Using muzzle pat- [52] S. Kumar, S.K. Singh, Feature selection and recognition of face by using hybrid
tern recognition as a biometric approach for cattle identification, Trans. ASABE 50 chaotic PSO-BFO and appearance-based recognition algorithms, Int. J. Nat.
(3) (2007) 1073–1080. Comput. Res. (IJNCR). 5 (3) (2015 Jul 1) 26–53.
[20] A.I. Awad, H.M. Zawbaa, H.A. Mahmoud, E.H.H.A. Nabi, R.H. Fayed, A.E. [53] S. Mishra, O.S. Tomer, E. Kalm, Muzzle dermatoglypics: a new method to identify
Hassanien, A robust cattle identification scheme using muzzle print images, in: bovines. Asian Livestock, FAO, 1995.
Proceedings of IEEE Federated Conference on Computer Science and Information [54] S. Kumar, S.K. Singh, A.K. Singh, Muzzle point pattern based techniques for in-
Systems (FedCSIS), 2013, pp. 529–534. dividual cattle identification, IET Image Proc. (2017 Jan 5).
[21] A. Noviyanto, A.M. Arymurthy, Beef cattle identification based on muzzle pattern [55] S. Kumar, D. Datta, S.K. Singh, A.K. Sangaiah, An intelligent decision computing
using a matching refinement technique in the sift method, Comp. Electr. Agr. 99 paradigm for crowd monitoring in the smart city, J. Parallel Distrib. Comput.
(2013) 77–84. (2017), http://dx.doi.org/10.1016/j.jpdc.2017.03.002.
[22] R. Tillett, C. Onyango, J. Marchant, Using model-based image processing to track [56] C.E. Bugge, J. Burkhardt, K.S. Dugstad, T.B. Enger, M. Kasprzycka, A. Kleinauskas,
animal movements, Comp. Electr. Agr. 17 (2) (1997) 249–261. M. Myhre, K. Scheffler, S. Ström, S. Vetlesen, Biometric methods of animal identi-
[23] Kumar S, Tiwari S., Singh S.K, Face recognition for cattle, in: Proceedings of 3rd fication, Course notes Laboratory Animal Science at the Norwegian School of
IEEE International Conference on Image Information Processing (ICIIP), 2015, pp. Veterinary Science, 2011, pp. 1–6.
65–72. [57] S. Ahmed, T. Gaber, A. Tharwat, A.E. Hassanien, V. Snáel, V., Muzzle-based cattle
[24] T. Gaber, A. Tharwat, A.E. Hassanien, V. Snasel, Biometric cattle identification identification using speed up robust feature approach, in: Proceeding of IEEE
approach based on webers local descriptor and adaboost classifier, Comp. Electr. International Conference on Intelligent Networking and Collaborative Systems
Agr. 122 (2016) 55–66. (INCOS) 2015, September, pp. 99–104, 2015.
[25] E.D. Pisano, S. Zong, B.M. Hemminger, M. DeLuca, R.E. Johnston, K. Muller, [58] W. Andrew, S. Hannuna, N. Campbell, T. Burghardt, Automatic individual holstein
M.P. Braeuning, S.M. Pizer, Contrast limited adaptive histogram equalization image friesian cattle identification via selective local coat pattern matching in RGB-D
processing to improve the detection of simulated spiculations in dense mammo- imagery, in: Proceedings of IEEE International Conference on Image Processing
grams, J. Digital Imaging 11 (4) (1998) 193–200. (ICIP) September, pp. 484–488, 2016.

16
S. Kumar et al. Measurement 116 (2018) 1–17

[59] Z. Kaixuan, H. Dongjian, Recognition of individual dairy cattle based on convolu- carabaos [in the Philippines]. Philippine Journal of Veterinary Medicine 1975
tional neural networks, Trans. Chinese Soc. Agr. Eng. 31 (5) (2015). (source: http://agris.fao.org/agris-search/search.do?recordID=PH19770204765).
[60] S. Mishra, V. Khune, D. Bhonsle, B.K. Sahu, A. Dubey, E. Kalm, Studies on muzzle [62] Y. Niu, Y. Sasaki, Recognition of muzzle pattern using local features in cattle,
pattern and inheritance of muzzle dermatoglyph in Sahiwal cattle. Bioimages 1 (2) (1993) 81–88.
[61] J.A. Solis, C.P. Maala, Muzzle printing as a method for identification of cattle and

17

View publication stats