b i o s y s t e m s e n g i n e e r i n g 1 8 8 ( 2 0 1 9 ) 2 2 9 e2 4 2

Available online at www.sciencedirect.com

ScienceDirect

journal homepage: www.elsevier.com/locate/issn/15375110

Research Paper

A machine vision system for early detection and

prediction of sick birds: A broiler chicken model

Cedric Okinda a, Mingzhou Lu a, Longshen Liu a, Innocent Nyalala a,

Caroline Muneri b,c, Jintao Wang a, Hailin Zhang a, Mingxia Shen a,*
a
College of Engineering, Laboratory of Modern Facility Agriculture Technology and Equipment Engineering of
Jiangsu Province, Nanjing Agricultural University, Jiangsu 210031, PR China
b
College of Veterinary Medicine, Nanjing Agricultural University, Nanjing, Jiangsu 210095, PR China
c
Faculty of Veterinary Medicine and Surgery, Department of Veterinary Surgery, Theriogenology and Medicine,
Egerton University, Njoro, Kenya

article info
The occurrence of poultry diseases not only affects farm production economics but also
Article history: leads to poor poultry welfare, food safety concerns, and zoonotic infections. Therefore,
Received 4 February 2019 timely detection of these diseases is of paramount importance in poultry production. This
Received in revised form study proposes a machine vision-based monitoring system for broiler chicken as they walk
31 August 2019 through a test area. Data were collected from two groups of broilers; control group and
Accepted 29 September 2019 treatment group (inoculated intramuscularly with virulent Newcastle disease virus)
housed in fully isolated chambers for comparative monitoring. The broilers were moni-
tored by video surveillance for data labelling and depth camera for the automated health
Keywords: status classifier development. Feature variables were extracted based on 2D posture shape
Poultry welfare descriptors (circle variance, elongation, convexity, complexity, and eccentricity) and
Shape descriptors mobility feature (walk speed). A statistical analysis of the feature variables established that
Posture all investigated features were statistically significant ðp < 0:05Þ with time after challenge in
Mobility the treatment group. The earliest possible infection detection time was on the 4th day
Support vector machine based on circle variance and elongation, and the 6th day based on eccentricity and walk
speed. However, convexity and complexity could not provide early detection. Two sets of
classifiers were then developed based on only the posture shape descriptors, and on all the
feature variables, The Support Vector Machine (RBF-SVM) outperformed all the other
models with an accuracy of 0.975 and 0.978 respectively. The proposed system can serve as
an automatic broiler monitoring system by providing an early warning and prediction of an
occurrence of disease continuously and non-intrusively.
© 2019 IAgrE. Published by Elsevier Ltd. All rights reserved.

* Corresponding author.
E-mail address: [email protected] (M. Shen).
https://doi.org/10.1016/j.biosystemseng.2019.09.015
1537-5110/© 2019 IAgrE. Published by Elsevier Ltd. All rights reserved.
230 b i o s y s t e m s e n g i n e e r i n g 1 8 8 ( 2 0 1 9 ) 2 2 9 e2 4 2

pandemics (Peiris et al., 2016; Rushton et al., 2005). Addition-

Nomenclature ally, according to Welfare Quality® (2009), animal health
condition is a key indicator of good welfare practice. This
2D Two dimensional
study focuses on broiler chicken, which is currently the main
3D Three dimensional
contributor to poultry meat production (OECD-FAO, 2017).
ANN Artificial Neural Network
o The conventional mode for chicken disease detection is by
C Degree Celsius
visual observations and sound distinction by farmers and
CPU Central Processing Unit
veterinarians (Huang et al., 2019; Zhuang et al., 2018; Zhuang &
fps Frames per Second
Zhang, 2019). However, in large-scale production, these
GB Gigabyte
methods of detection are time-consuming, subjective, labour-
GC Control group
intensive, and fail to provide an early detection (Zhuang et al.,
GT Treatment group
2018). Tablante (2013) and Butcher, Jacob, and Mather (1999)
GHz Gigahertz
reported on common poultry diseases, visible signs of each
HD High Definition
infection, and the modes of prevention and control. Addi-
IQR Interquartile Range
tionally, the Chicken Health Handbook by Damerow (2016)
MAT Medial Axis Transform
described how to identify sick birds by several methodolo-
mdl1 Model 1, based on only the shape descriptors
gies based on observations such as behavioural changes,
mdl2 Model 2, based on both the shape descriptors
depressed bird look (posture), variation in egg production,
and walk speed
comb appearance, droppings examination, and post-mortem
MOV Movie Digital Video
examination.
NDV Newcastle Disease Virus
Our study presents a posture and mobility evaluation to
p Probability value (p-value)
provide a technique for distinguishing between healthy and
PC Personal Computer
sick birds. Generally, infected poultry always have a distinctly
PLF Precision Livestock Farming
different body posture compared to healthy poultry (Butcher
Рt Extracted feature variables
et al., 1999; Damerow, 2016; Manning et al., 2007; Zhuang
RBF Radial basis function
et al., 2018). Similarly, mobility is another important factor
ROI Region of Interest
that has been used to examine the health condition of birds. A
St 2D posture shape descriptors
sick bird is often reluctant to walk for very long and will isolate
SVM Support Vector Machine
itself and displays a depressed bird look/posture (Damerow,
SDK Software Development Kit
2016). Furthermore, Paul-Murphy and Hawkins (2014) re-
USB Universal Serial Bus
ported that a bird in pain has a decreased social interaction
V2 Version 2
and will always seclude itself and remain stationary at a dis-
gc Circle Variance
tance from the rest of the flock.
gE Shape Elongation
Several studies have reported different techniques based
gcv Shape Convexity
on machine vision to monitor various bio-responses and
gcx Shape Complexity
bioprocesses that indicate health conditions in broilers such
ge Shape Eccentricity
as weight-gain, behaviours, mobility, posture, and activities.
gs Chicken walk speed
Weight-gain is vital information in livestock production. It is
an indicator of body growth, feed conversion ratio, and market
readiness (Wongsriworaphon, Arnonkijpanich, &
Pathumnakul, 2015). The first two are indicators of good
1. Introduction health and good animal welfare, while the latter is applied in
supply chain management. Additionally, good animal welfare
There has been a rise in the consumption of poultry meat and is characterised by good health and productivity. Thus,
poultry meat products in the recent years, with a projected monitoring animal weight gain can be used to detect disease
increase within the next decade (Henchion, McCarthy, occurrences and other vitality issues for necessary counter-
Resconi, & Troy, 2014). Consequently, global poultry produc- measures to be undertaken (Menesatti et al., 2014;
tion has significantly increased, with a production of about Mortensen, Lisouski, & Ahrendt, 2016). Generally, mobility is
117 Mt in 2016 (OECD-FAO, 2017). With the large-scale pro- always associated with walking or locomotion as it describes
duction and breeding of livestock, the production risks the quality or state of being mobile or ability to move. In
involved have also increased (Manning, Chadd, & Baines, 2007; poultry, immobility is often a sign of birds experiencing some
Sims, 2008). The main challenge affecting poultry production discomfort (Paul-Murphy & Hawkins, 2014). These discom-
is the frequent occurrence of poultry diseases (Huang, Wang, forts may be as a result of skeleton (leg) disorders (Bradshaw,
& Zhang, 2019; Zhuang, Bi, Guo, Wu, & Zhang, 2018; Zhuang Kirkden, & Broom, 2002), nutrition deficiencies (Bradshaw
& Zhang, 2019), which has often led to significant losses in a et al., 2002), and health conditions such as leg health and
flock or loss of entire flocks (Rushton, Viscarra, Bleich, & diseases (Butcher et al., 1999; Tablante, 2013). Additionally,
McLeod, 2005). Furthermore, poultry diseases pose a severe Aydin (2017a) reported that sound birds walk significantly
threat to food safety if infected meat and products are inges- faster than unsound birds. Thus, an infection greatly affects
ted (Friel & Ford, 2015). Poultry diseases like the avian influ- the mobility characteristics of a bird. Chicken mobility and
enza virus are zoonotic and have led to several global activeness have been used to assess the degree of lameness in
 b i o s y s t e m s e n g i n e e r i n g 1 8 8 ( 2 0 1 9 ) 2 2 9 e2 4 2 231

several studies (Aydin, 2017a; 2017b). Furthermore, lameness old Arbor Acres broilers were obtained from the farm's
in broilers describes a range of injuries resulting from an hatchery and were reared in fully isolated controlled envi-
infection or non-infective origin (Thorp & Duff, 1988). Aydin ronment chambers. The temperature was set at 34  C at the
(2017b) established that inactivity is highly correlated to start of the experiment and gradually decreased until 26  C on
lameness. Furthermore, Aydin (2017a) reported that broiler the last day of the experiment. The humidity was set at 50% at
walk speed, step frequency, step length, and body oscillations the beginning of the experiment and was gradually increased
are statistically correlated to the level of lameness. Kristensen to 80% on the final day of the experiment. The chamber floor
and Cornou (2011) developed an automated system that could was litter (50% sphagnum and 50% wood-shavings). Feeding
identify deviations in activity levels in a group of broiler was performed according to the study by Aydin (2017b), and
chicken. The study characterised the undisturbed activity anti-stress vitamins were also given at strategic periods.
levels by age and could provide notification in any activity Water was given ad libitum during the experimental period.
deviation to detect potential problems in the flock. The lighting regime consisted of two dark periods, following
As mentioned earlier, depressed-bird-look/posture obser- the study by Mortensen et al. (2016), with the light periods
vation is a popular technique to identify sick birds. Zhuang having an approximate light intensity of 30 lx for the duration
et al. (2018) presented a system that could analyse broiler of the experiment.
posture for early detection of sick broilers based on chicken In each chamber, a wooden test corridor of dimensions
image skeleton features for images taken from the side. 2.50 m (length) by 0.50 m (width) by 0.50 m (height), a 3D Kinect
Additionally, Zhuang and Zhang (2019) developed a sick camera for Windows V2 (Microsoft Corp., Washington, USA)
broiler detection system based on deep learning technique and a HD video surveillance camera DS-2CD3T35-13 (HIKVI-
using the B1 broiler and PASCAL VOC2012 databases. Both SION) were pre-installed before the chickens were brought
systems successfully classified sick and healthy broilers. into the chambers. The Kinect camera was positioned 2.50 m
Aerial posture analysis has also been presented by Pereira above the ground and precisely at the centre of the testing
et al. (2013) to assess broiler welfare condition by use of an corridor, as shown in Fig. 1. The Kinect camera was connected
overhead video camera to develop a comprehensive assess- to an Intel core i5-4500U CPU, 4 GHz, 16 GB physical memory
ment system for broiler behavioural expressions. (Intel, Santa Clara, CA, USA), Microsoft Windows 10 PC
Moving on from earlier studies that have been performed installed with the Kinect for Windows Software Development
in disease detection in broilers by image processing tech- Kit (SDK) via a USB port. Depth images (512  424 pixels) were
niques, for real farm implementation, overhead images would acquired from the Kinect camera at 1 fps using MATLAB
be most preferred due to the ease of camera positioning in R2018a (The MathWorks Inc., Natick, MA) image acquisition
large-scale farms and the least disturbance to the animals toolkit. An infrared (IR) depth sensor was applied due to its
(Van der Stuyft, Schofield, Randall, Wambacq, & Goedseels, invariance to variations in ambient light conditions (Kongsro,
1991). Additionally, overhead images pose a less arduous 2014; Okinda, Lu, Nyalala, Li, & Shen, 2018). Additionally, the
task of background removal. Currently, there is no literature surveillance camera was used to acquire videos (for observa-
on the analysis of overhead posture shape to detect disease tion, labelling, and verification) at 22 fps MOV format. Figure 2
occurrence in broilers. Thus, this study proposes a method- shows clear broiler colour images for the top view and side
ology based on a non-intrusive depth camera sensor and view during the experiment.
machine vision for early detection and prediction of sick At 21 days old, the birds were tagged (poultry feet tags,
broilers. Geshifeng Company) and randomly divided into two groups of
The primary objective of this study was to develop an auto- 20 birds each; a total of 280 broilers were used in this study.
classification system that could discriminate between healthy The first group was a treatment group ðGT Þ and was inoculated
and sick broilers. The specific objectives included developing: intramuscularly with 200 ml virulent Newcastle Disease Virus
an efficient depth image processing algorithm, an effective
and robust posture analysis algorithm for 2D shape geometric
features extraction, an algorithm to compute the broiler walk
speed, a classifier to establish a correlation of broiler feature
variables to the health status, and the statistical relationships
between the extracted feature variables with an infection
exposure period. This proposed system could serve as a Pre-
cision Livestock Farming (PLF) system if applied in a practical
farm scenario by providing support to the stockmen in the
detection of avian diseases in broiler production.

2. Materials and methods

2.1. Experiment design and monitoring system setup

Experiments were conducted between January 2018 and

March 2019 at Ping Du Poultry Farm of New Hope Liu He
Limited Company, Qingdao, Shandong Province, China. Day- Fig. 1 e Experiment setup and image acquisition system.
232 b i o s y s t e m s e n g i n e e r i n g 1 8 8 ( 2 0 1 9 ) 2 2 9 e2 4 2

Fig. 2 e Colour images of broiler chicken (a) top view and (b) side view during the experiment. (For interpretation of the
references to color in this figure legend, the reader is referred to the Web version of this article).

(NDV) with a titre of 106 EID50 per bird. The NDV was obtained posture shape descriptors ðSt Þ. Comparison with the video
from Nanjing Agricultural University, College of Animal observation was used to validate the proposed health classifier.
Research and was propagated in 10-day-old specific- Additionally, the effect of disease progression (time in days after
pathogen-free (SPF) embryonated chicken eggs (ECEs), challenge) on the extracted features was analysed to establish
following the study of Mast et al. (2006). The second group was the earliest possible time an infection manifested itself through
used as the control group ðGC Þ: In each experiment, a bird was the extracted feature variables ðРt Þ, where gs 2Рt and St 2Рt .
placed at the starting point of the testing corridor, and over- Feature extraction was performed by image analysis on the
head depth images of the experimental setup were captured processed depth images. Health classifier models were then
while the broiler walked from the starting point to the end of developed, to determine the model performance difference
the testing corridor. This procedure was repeated for all the between mobility feature gs and image features St . Two sets of
chickens in each chamber daily. To prevent the chicken from models were developed; one based solely on St ðmdl1 Þ while
ground pecking and to attract them to walk through the test the other was a combination of all the feature variables ðmdl2 Þ.
corridor, the corridor's floor was kept clean at all times, and The algorithm flow diagram of the proposed system is pre-
some feed and a few chickens were placed at the opposite end sented in Fig. 3.
of the test corridor. Thus, the broiler under test would walk
through to join them due to social attraction to cluster 2.3.1. Image pre-processing algorithm
together (Febrer, Jones, Donnelly, & Dawkins, 2006). A raw depth image data was processed as follows. Firstly,
During the experiment period, overhead images were ac- background removal by image subtraction technique was
quired, and clinical symptoms observed for all the chicken for performed to remove the floor and test corridor according to
14 days (35 days old) with Day 0 being the NDV inoculation Eq. (1).
day. All experiments were carried out in Nanjing Agricultural
 
0 if Fx;y  Bx;y   T
University facilities, in compliance with and using protocols Gx;y ¼ ; (1)
Fx;y otherwise
(protocol JX524203/1) approved by the Biosafety Committee of
Nanjing Agricultural University. The handling of the broiler where, Gx;y is the resultant image with background removed,
chicken was performed according to the guidelines approved Fx;y is the original image, Bx;y is the background image, and T is
by the experimental animal administration and ethics com- the threshold. Secondly, based on distance intensities of the
mittee of Nanjing Agricultural University. depth image (Jana, 2012), maximum and minimum thresholds
were set to remain within the Region of Interest (ROI) as pre-
2.2. Data labelling sented in Eq. (2).


All the acquired depth data for both sick and healthy birds Gx;y if Tmn  Gx;y  Tmx
Hx;y ¼ ; (2)
0 otherwise
were manually labelled by surveillance video observation. The
timestamps of the moment a broiler entered and exited the where, Hx;y is the resultant depth image after depth threshold,
test corridor were also labelled for chicken walk speed Tmn is the minimum depth distance threshold, and Tmx is the
extraction. maximum depth distance threshold. Thirdly, Hx;y was
smoothed by 5  5 pixels zero-mean Gaussian kernel filter (Eq.
2.3. Image pre-processing and feature extraction (3)) then morphological opening (Eq. (4)) was performed by a
algorithms circular structural element of size 12 pixels to remove local
minima associated with both head and body and to obtain a
These algorithms aimed at detecting the health condition of a clear depth image. Finally, the image was then converted to a
broiler chicken were based on the walk-speed ðgs Þ and simple 2D binary image by Otsu's method (Otsu, 1979).
 b i o s y s t e m s e n g i n e e r i n g 1 8 8 ( 2 0 1 9 ) 2 2 9 e2 4 2 233

Fig. 3 e The algorithm flow diagram of the proposed system. (a) acquired depth image, (b) segmented broiler image, (c)
extracted feature variables, and (d) developed models.

P qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pk1
  a circle. The terms mP ¼ 1k k1
i¼1 ri and sP ¼
1
k i¼1 ðri  mP Þ are
2

Ix;y ¼ Hx;y 5u (3)

the mean and standard deviation of radial distances from the
where, Ix;y ¼ Iðx; yÞ ¼ I is the resultant filtered image, and u ¼ centre of gravity (g) to the shape boundary points (bp) of the 2D
0  1 shape respectively, ri ¼ jjbpi  g|| is the radial distance of the ith
s2 þt2
B 1 12 C bp from g and k is the total number of bp. Shape Elongation ðgE Þ
uðs; tÞ ¼ B C is the Gaussian filter kernel.
s2

@2ps2 e A
(Eq. (6)) is a parameter based on a shape's minimum bounding
box, it represents the degree of stretch, where RL and RW are
 the length and width of the minimum bounding rectangle,
Ax;y ¼ I , Se ¼ ðI 4 SeÞqSe ¼ ∪ ðSeÞz jðSeÞz (4)
respectively. Shape convexity ðgcv Þ (Eq. (7)) is a parameter
where, Ax;y is the resultant opened image, Se is the structural related to the shape convex hull in terms of the perimeter. It
element, ðSeÞz is the translation of Se by a point z. presents the degree of shape convexity in an epigraph, where
PH is the convex hull perimeter and PS is the 2D shape
2.3.2. Feature extraction algorithm perimeter.
From previous studies on the application of computer vision The measure of shape complexity ðgcx Þ (Eq. (8)) was based
in livestock image analysis, several image-related features on the study by Panagiotakis and Argyros (2016), whereby
have been proposed and applied to develop correlations to a shape complexity is defined as a function of the entropy of the
targeted bio-response or a bio-process (Mollah, Hasan, Salam, extracted Medial Axis Transform (MAT) 16-bin histograms,
& Ali, 2010; Mortensen et al., 2016). However, the proposed where jWj is the number of edges contained in the shape's
technique approaches image analysis from a shape repre- MAT graph, pij is the jth bin of the ith histogram and S is the
sentation point of view by extracting St (Zhang & Lu, 2004) also skeleton hence logjSj represents the global information of the
known as shape geometric parameters (Kurnianggoro & Jo, skeleton. The MAT is a locus of centroids and their radii
2018) for correlation to the broiler health status. For each describing maximal circles contained in a shape and are
bird, ten images were selected for feature extraction. The tangent to the shape's boundary points (Zhang & Lu, 2004) as
equations for the extraction of Рt are summarised in Table 1, shown in Fig. 4 (c) and (d). Eccentricity ðge Þ (Eq. (9)) is the ratio
while the extraction of St from a processed depth image is of the Eigenvalues (l1 and l2) of a covariance matrix of a fitted
presented in Fig. 4. ellipse over a 2D shape, and these Eigenvalues correspond to
Circle variance also known as circularity ðgc Þ (Eq. (5)) is a the ellipse's major and the minor axis respectively. Chicken
parameter that presents the degree of similarity of a shape to walk speed ðgS Þ (Eq. (10)) is the average speed at which the
broiler walked from the start to the end of the test corridor.
Since the depth data were acquired at 1 fps, tk was equal to the
Table 1 e The extracted features variables. number of frames captured for the kth chicken, and l being the
Extracted Features Defining Equations length of the test corridor.

Circle variance sP
gc ¼ ð5Þ 2.4. Health status prediction algorithm
mP
Shape elongation RL
gE ¼ ð6Þ
RW To predict the health status of the broilers, six prediction
Shape convexity PH
gcv ¼ ð7Þ models were explored. The models were developed based on
PS
jWj Рt . The summary of the dataset used in this study is given in
Shape complexity PP 16
gcx ¼  ðpij logpij þlogjSjÞ (8) Table 2. The dataset was selected based on the criteria that the
i¼1 j¼1

Shape eccentricity l1 entire broiler image must be within the screen field and does
ge ¼ ð9Þ
l2 not touch any of the image borders. A total of 34,280 images
Chicken walk speed (m s1) l were used in this study. The training set was 70% of the total
gs ¼ ð10Þ
tk dataset while the remaining 30% as the testing set.
234 b i o s y s t e m s e n g i n e e r i n g 1 8 8 ( 2 0 1 9 ) 2 2 9 e2 4 2

Fig. 4 e The extraction of 2D broiler posture shape parameters: (a) and (b) extraction of circle variance (white circle), shape
elongation (yellow rectangle), and eccentricity (green ellipse) on a 2D broiler posture shapes of a broiler in the control group
ðGC Þ and treatment group ðGT Þ, respectively. (c) and (d) extraction of shape complexity on a 2D broiler posture shapes of a
broiler in the control group ðGC Þ and treatment group ðGT Þ respectively.

Table 2 e Summary of modelling dataset. score between related groups to examine when the differ-
ences occur on subsequent group combinations. Note that in
Dataset Number of images
this study, the data was not normally distributed even after
Training dataset 23,996
data transformation. Hence, nonparametric statistical tech-
Testing dataset 10,284
niques were applied. All statistical computations were per-
Total dataset 34,280
formed by Statistical Package for the Social Sciences (IBM SPSS
Statistics).
The explored models were support vector machine (SVM) Based on the quartile values of Рt in GC , and the direction of
(linear, quadratic, cubic and radial basis function (RBF)), arti- association of Рt in GT with time after challenge, a threshold
ficial neural network (ANN), and logit regression. These ðdi Þ given by Eq. (11) was set on Рt for GT between Day 3 to Day 7
models were trained by a 10-fold cross-validation-based for early disease detection in GT i.e., for a þve rsi , Рt values
parameter search on the training dataset after which they greater than di indicates an infection, while for a -ve rsi , Рt
were evaluated on a testing dataset. A comparison to the values less than di indicates an infection.
manual labelling of the testing dataset was presented in terms

MaxРt i if rsi is þ ve
of prediction accuracy. di ¼ i ¼ 1; 2; …; N (11)
MinРt i otherwise

2.5. Early disease detection where di is the threshold value for the ith feature variable in Рt ,
N is the total number of feature variables in Рt ðN ¼ 6Þ, MaxРt i
For early disease detection, the developed system aimed at and MinРt i are the 3rd quartile and 1st quartile value of the ith
establishing the point in time at which infection becomes feature variable in Рt for GC as given in Table 7, and rsi is the
noticeable by analysing the variations or changes in Рt before coefficient of correlation of the ith Рt and time (computed from
any death occurs in GT with reference to GC . Firstly, a Spear- Spearman's RankeOrder correlation test). Note that the
man's RankeOrder correlation test was performed to establish quartiles were selected as the threshold values as a precaution
the direction (positive or negative) of the association between against outliers in the minimumemaximum margins.
Рt and time after challenge in GT . Secondly, the Friedman test
(Friedman, 1937) also known as Friedman's ANOVA was used
to analyse the effect of time on Рt in GT (grouped into days 3. Results
after challenge). The Friedman test is a non-parametric sta-
tistical analysis for variance that avoids the assumption of During the 14 days of observation and data collection after
data normality by comparing columns without the effects of NDV challenge, 99.29% of all the chickens in GT died, as shown
the rows. Hence, size and dependencies do not affect the test in Table 3. The clinical signs observed on the infected birds
results in the test sample (Aydin, 2017a, 2017b; Friedman, included respiratory distress, paralysis, circling, and torti-
1937). Finally, the Wilcoxon Signed-Rank Test (Taheri & collis. Manual verification by surveillance video observation
Hesamian, 2013) was applied to determine the statistical dif- established that the image acquisition system correctly
ferences among the groups (time in days after challenge captured the depth images at 1 fps as the broilers walked
groupings). This post hoc test investigated the changes in rank through the test corridor.
 b i o s y s t e m s e n g i n e e r i n g 1 8 8 ( 2 0 1 9 ) 2 2 9 e2 4 2 235

often fitted using a likelihood function, which is a max-

Table 3 e Summary of the performance of the treatment
imisation problem (General Maximum Likelihood (GML)).
group ðGT Þ after NDV challenge.
Therefore, the estimation of this maximum likelihood signif-
Day after Number of Broilers
icantly affects the overall model performance (Kleinbaum &
exposure
Alive Able to walk Unable to Dead Klein, 2010; Mansournia, Geroldinger, Greenland, & Heinze,
walk 2017). The SVM kernel functions and the logit log(likelihood)
0 140 140 0 0 parameters were computed iteratively during the 10-fold-
1 140 140 0 0 validation in the model training phase (parameters resulting
2 140 140 0 0 to highest validation accuracy while effectively avoiding
3 140 140 0 0
overfitting and underfitting of a model and having a good
4 140 140 0 0
5 140 140 0 0
generalisation ability). Similarly, for the ANN model, the
6 140 138 2 0 number of neurons in the hidden layers was set to 12 after
7 138 134 4 2 evaluation of the percent error during model validation.
8 124 99 25 14 However, as the number of neurons increases, percentage
9 105 80 25 19 error decreases at the expense of the computation time
10 77 45 32 28
(Wongsriworaphon et al., 2015). Therefore, in this current
11 36 7 29 41
study, the number of hidden layers was set to 12 to balance
12 6 2 4 30
13 2 0 2 4 the computation time and percentage error while training was
performed by scaled conjugate gradient backpropagation al-
gorithm (Møller, 1993).
3.1. Shape geometric features and mobility feature All models were evaluated on the testing dataset. Table 5
evaluation presents the performance of the models on the testing data-
set in terms of accuracy. It was established that RBF-SVM had
The Рt were extracted from all the selected images for both the best classification results while Logit regression returned
sick and healthy birds. The trend of Рt in terms of the mean the lowest accuracy irrespective of the training features.
and range between minimum and maximum (error) values There was no significant performance difference between
are presented in Fig. 5 for the entire observation period. mdl1 and mdl2 . However, mdl2 was slightly more accurate than
Generally, it was observed that gc and gcv increased while gE , mdl1 for all the developed models.
gcx , ge , and gs decreased with infection progression (exposure Additionally, the performance of each model was evalu-
time). From Fig. 5 the sharp rise or drop in the values of GT ated based on the relative error of daily testing dataset, as
features as from around the 6th to the 8th day represents the shown in Fig. 7. It was established that all the models had
point at which the infection starts to be manifested by a wide similar performance except for the cubic SVM in mdl1 .
variation in Рt . Furthermore, it was observed that the cubic SVM had the
At the start of the experiment, ground pecking and largest variations in its accuracy when mdl1 and mdl2 perfor-
preening behaviours by the birds were observed based on the mances were compared on the entire testing dataset as pre-
ethogram presented by Pereira et al. (2013) and Kristensen sented in Table 5.
et al. (2007). However, these behaviours were minimised in
the preceding experiments by increasing the number of 3.3. Early disease detection by statistical analysis
chickens used for social attraction as a motivation for the
broiler under test to cluster with the others (Aydin, 2017b; A correlation analysis established that there was a statistically
Febrer et al., 2006). However, as shown in Fig. 6, a distinction significant ðp < 0:01Þ positive correlation between both gc and
between healthy birds in GC ground pecking and infected birds gcv with time in days after challenge with coefficients
in GT can be characterised by the tail positioning, whereby in rs ¼ 0:883 and 0:596; respectively. However, it was established
the former during ground pecking activity the tail is usually that there was a statistically significant ðp < 0:01Þ negative
not dropped. Furthermore, several images of a single bird correlation between gE , gcx , ge , and gs with time after chal-
were acquired as it walked through the test corridor, thus, lenge, with coefficients rs ¼ 0:872; 0:695; 0:757; and 0:838
minimised the errors associated with the mismatch between respectively. This correlation test result concurs with the vi-
ground pecking and sick broilers in the classification. sual observations made on the trends of Рt as presented in
Fig. 5. Table 6 and Fig. 8 presents the statistical analysis of Рt
3.2. Health status classifier evaluation for the GT while Table 7 presents the descriptive statistics of Рt
for GC for the entire experiment period. The statistical analysis
The model parameters for all the trained models developed in for GT was performed as from Day 3 till Day 7 when the first
this study are presented in Table 4. Classifier parameters are broiler death occurred to eliminate the effects of the NDV
important factors that affect the overall performance of a inoculation procedure on the GT broilers during Day 0 to Day 2
model in terms of efficiency and accuracy. The choice of the (effect of injection), while Day 7 was included for comparison
kernel function is a critical factor in performance of an SVM purposes.
model (Zhuang et al., 2018). The commonly applied kernel By Friedman's test, there was a statistically significant
functions are the linear, polynomial (quadratic and cubic) and difference ðp < 0:05Þ in all the Рt depending on the time after
the RBF kernel functions. The logit regression predicts prob- NDV challenge. A post hoc test on combinations of Day 3 to Day
abilities rather than grouping data into classes. Therefore, it's 4, Day 4 to Day 5, and Day 5 to Day 6 was performed to
236 b i o s y s t e m s e n g i n e e r i n g 1 8 8 ( 2 0 1 9 ) 2 2 9 e2 4 2

Fig. 5 e The mean, maximum and minimum daily trend of the extracted feature variables during the observation period for
both the control group ðGC Þ and treatment group ðGT Þ broilers.

examine where the differences occur. Based on the study by terms of gc mean values. The mean values of gE on Days 4e7
Glickman, Rao, and Schultz (2014), it was not necessary to were significantly ðp < 0:05Þ lower than the gE mean for Day 3.
perform Bonferroni adjustment to prevent a deleterious effect Additionally, there was no statistically significant ðp < 0:05Þ
on the statistical inference (Perneger, 1998). differences between Days 4, 5 and 7 in terms of gE mean
According to Table 6, the mean values of gc on Days 4e7 values. The mean values of gcv on Day 7 was significantly
were significantly ðp < 0:05Þ greater than the gc mean for Day ðp < 0:05Þ greater than the gcv mean for Days 3e6. However,
3. However, there were no statistically significant ðp < 0:05Þ there was no statistically significant ðp < 0:05Þ differences
differences between Days 4 and 5 nor between Days 6 and 7 in between Days 3, 4 and 5 in terms of gcv mean values. The mean
 b i o s y s t e m s e n g i n e e r i n g 1 8 8 ( 2 0 1 9 ) 2 2 9 e2 4 2 237

Figure 8 shows the early infection detection based on the

above threshold on GT . It was established that only gc , gE , ge
and gs could provide an early detection on Days 4, 4, 6 and 6
respectively. However, early detection could not be achieved
by gcv and gcx before any death occurs in GT .

4. Discussion

NDV is one of the common poultry diseases (Butcher et al.,

1999) caused by NDV Avulavirus (Waterson, Pennington, &
Allan, 1967). The housing chambers used during the experi-
ment were completely insulated to prevent inter-group
infection because NDV can be propagated by air from one
bird to another (Chansiripornchai & Sasipreeyajan, 2006).
From Table 3, the severe effects of the infection were apparent
from Day 9, but substantial effects in posture and mobility
Fig. 6 e An illustration of the differences in the 2D posture were visible by Day 7. This was due to NDV's short incubation
shape on Day 7 for (a and b) a control group (GC Þ bird in a period of 2e15 days (Rahman, Rabbani, Uddin, Chakrabartty,
walking position and in a ground pecking position & Her, 2012).
respectively and (c) a treatment group ðGT Þ bird in a The proposed system introduces an automated monitoring
walking position. technique of chicken welfare-related bio-responses by a ma-
chine vision system. Image preprocessing algorithm was
developed as the initial step in the visual diagnosis of broiler
Table 4 e The parameters of the developed recognition disease (Zhuang et al., 2018) to segment the chicken shape
models. from the background. Secondly, St were extracted by image
Classifier parameters Mdl1 Mdl2 analysis algorithm. St have description ability and can be used
to differentiate shapes in shape analysis, but to achieve a
SVM Kernel parameters Scale degree Scale degree
higher discriminative power several St are always combined
Linear 0.4083 e 0.4877 e (Kurnianggoro & Jo, 2018; Zhang & Lu, 2004). Therefore, in this
Quadratic 0.3871 2 0.4262 2
proposed technique, St was described by five geometric pa-
Cubic 0.3360 3 0.5041 3
rameters to improve the descriptive ability.
RBF 0.5600 e 0.6100 e
Logit log(likelihood) 9273.413 8774.085 Consider a comparison of GT to GC in Fig. 5, the gc and gcv
ANN topology 5e12 e 2 6e12 e 2 were observed to increase with exposure time from Day 4 and
Day 7 respectively, from the observation that an infected
chicken displayed a typical sick bird posture-hunched stance,
values of gcx on Days 4e7 were significantly ðp < 0:05Þ lower characterised by a tucked-in neck and dropped tail. Hence, the
than the gcx mean for Day 3. However, there were no statis- shape of a sick broiler appeared more circular and more convex
tically significant ðp < 0:05Þ differences between Days 4 and 5 compared to a healthy broiler with exposure time, as shown in
in terms of gcx mean values. The mean values of ge on Days Fig. 4 and Fig. 9. The gE , gcx , and ge were observed to decrease
4e7 were significantly ðp < 0:05Þ lower than the ge mean for with exposure time as from Day 4, Day 7 and Day 8 respectively.
Day 3. Furthermore, there was no statistically significant gE is minimum bounding box dependent, a healthy bird's
ðp < 0:05Þ differences between Days 4 and 5 nor between Days bounding box has a longer longitudinal length compared to a
6 and 7 in terms of ge mean values. The mean values of gs on sick broiler. Hence, the degree of stretch is higher than that of a
Days 6e7 were significantly ðp < 0:05Þ lower than the gs mean sick bird. The gcx adopted in this study is MAT dependent, the
for Days 3e5. However, there were no statistically significant more circular a 2D shape is, the fewer the maximal inscribed
ðp < 0:05Þ differences between Days 3 and 4 nor between Days circles, hence, low shape complexity in sick birds (Fig. 4 c and d).
6 and 7 in terms of gs mean values. Finally, ge is a parameter dependent on an inscribed ellipse on a
shape it is the extent to which the fitted ellipse departs from
circularity, for a circular shape, ge is zero. Thus, sick birds will
have a lower eccentricity because their shape has a higher
Table 5 e Comparison of the performance of different circularity compared to the healthy broiler. Figure 9 presents a
models on the testing dataset (accuracy). comparison of two chickens from each group with their images
Models mdl1 mdl2 on Days 0, 4, 7, 10, and 13. A clear visual distinction can be
Linear SVM 0.858 0.860 observed as from Day 10, whereby the bird in GT appears more
Quadratic SVM 0.961 0.965 circular and less elongated than the bird in GC .
Cubic SVM 0.917 0.971 The chicken mobility in terms of gs in GT was observed to
RBF SVM 0.975 0.978 decrease from an average of 0.173 m s1 on the day of inocu-
ANN 0.969 0.969 lation (Day 0) to complete immobility with disease progres-
Logit 0.817 0.808
sion. Additionally, the average walk speed of GC was
238 b i o s y s t e m s e n g i n e e r i n g 1 8 8 ( 2 0 1 9 ) 2 2 9 e2 4 2

Fig. 7 e The performance of the developed health classifiers on the daily testing dataset.

Table 6 e The statistical results of the Рt of birds in the treatment group ðGT Þ expressed in mean values for Day 3 to Day 7
after NDV challenge.
Days after exposure Circularity Elongation Convexity Complexity Eccentricity Walk speed (ms1)
ðgc Þ mean ðgE Þ mean ðgcv Þ mean ðgcx Þ mean ðge Þ mean ðgs Þ mean
Day 3 0.585c 2.665a 0.991b 10.975a 0.924a 0.157b
Day 4 0.646b 2.256b 0.993b 10.776c 0.893b 0.156b
Day 5 0.647b 2.184b 0.990b 10.701c 0.893b 0.161a
Day 6 0.730a 2.052c 0.986c 10.834b 0.876c 0.114c
Day 7 0.737a 2.195b 0.997a 10.081d 0.872c 0.111c

a, b, c, d Mean ranks within a column, with no superscript in common differ significantly ðp < 0:05Þ.

0.164 m s1 for the entire experiment period as shown in Table In an efficient and effective disease detection system in
7. This finding concurs with those of Aydin (2017a), Na €a
€ s et al. poultry production, the occurrence of any infection should be
(2010), and Caplen et al. (2012), that healthy sound birds walk detected before the death of any bird in a flock. Based on the
significantly faster between a range of 0.15 m s1 to 0.17 m s1, correlation test results, early disease detection could be ach-
while unsound birds walked at a lower speed of below ieved by establishing the time at which gc and gcv starts to
0.11 m s1. Additionally, the results obtained in our study increase, while gE , gcx , ge , and gs start to decrease before the
concur with these previous reports based on the statistical first broiler death. However, it was established that gcv , and gcx
analysis of gs for early disease detection, as presented in Table could not provide an early detection based on the set
6, whereby Day 6 had a mean gs of 0.114 m s1 and it is the threshold technique. Furthermore, from the correlation re-
initial point in time at which some of the broilers were sults gcv and gcx had weak correlation strength with time in
immobile as seen in Table 3. days after exposure compared to the other feature variables.
A comparison of the developed models presented in Table gc and gE could provide an early detection on Day 4, while ge
5 shows that RBF-SVM outperformed all the other models. and gs on Day 6: Moreover, from Fig. 8 it can be observed that
Additionally, it had a good generalisation ability and robust- gc (mean and median (IQR)) increases in Day 4 when compared
ness on the testing dataset as presented in Fig. 7. However, a to Day 3. Similarly, gE (mean and median (IQR)) decreases in
high prediction error occurred on the first two days after NDV Day 4 when compared to Day 3. Additionally, both gs and ge
challenge (Fig. 7), because the NDV was still in its incubation (mean and median (IQR)) are also observed to decrease in Day
phase. Hence, the birds in GT had a similar appearance to 6 when compared to Day 5. Furthermore, according to Aydin
those in GC . Thus, infected birds were predicted to belong to (2017a) the gs (mean) for Day 3e5 are considered as gs of
GC . Moreover, this is why Day 0 to Day 2 was not considered in sound birds while gs (mean) for Day 6 would be considered as
the statistical analysis for early disease detection in Table 6. A gs of unsound birds. It was also noted that the infection could
comparison of mdl1 to mdl2 showed that there was no signifi- only be detected after the incubation period of the NDV Avu-
cant difference between the models in terms of model accu- lavirus. The Рt of some days after challenge (Day 4 and 5 for gc
racy. Hence, it can be concluded that St can be applied as and gE , Day 3, 4, 5 and 6 for gcv , and Day 3 and 4 for gs ) were
predictors in broiler disease prediction. Furthermore, gs has quite similar as seen in Table 6. Therefore, accurate detection
been applied in several studies to access the gait score (Aydin, of variation of feature variables on these days would be quite
2017a; Caplen et al., 2012), these studies were based on difficult. Hence, for precise identification of the effect of an
lameness detection and established that sound birds walk infection on Рt on these days, a new feature should be
significantly faster. Thus, walk speed as a mobility feature can included in the proposed system or the system should be in-
be used as a feature variable in disease detection and predic- tegrated with other welfare, bio-response related monitoring
tion systems in birds. systems such as weight estimation or activity detection.
 b i o s y s t e m s e n g i n e e r i n g 1 8 8 ( 2 0 1 9 ) 2 2 9 e2 4 2 239

Fig. 8 e The box plots presenting the statistical analysis for the extracted feature variables for Day 3 till Day 7, the thresholds
for infection detection and the detected infection days in the treatment group ðGT Þ.
240 b i o s y s t e m s e n g i n e e r i n g 1 8 8 ( 2 0 1 9 ) 2 2 9 e2 4 2

Table 7 e The descriptive statistic for the Рt of birds in the control group ðGC Þ for the entire experiment period (n ¼ 19,580).
Feature variables sCircularity Elongation Convexity Complexity Eccentricity Walk speed (ms1)
ðРt Þ ðgc Þ ðgE Þ ðgcv Þ ðgcx Þ ðge Þ ðgs Þ
Minimum 0.520 2.400 0.983 9.994 0.850 0.125
Maximum 0.690 2.802 0.999 11.200 0.940 0.192
Mean 0.595 2.639 0.991 10.760 0.905 0.164
1st Quartile 0.563 2.592 0.988 10.606 0.890 0.156
3rd Quartile 0.626 2.711 0.993 10.900 0.922 0.179
Median 0.5911 2.652 0.991 10.751 0.908 0.167
Standard deviation 4.100  102 9.800  102 3.505  103 2.036  101 2.140  102 1.992  102
Variance 1.679  103 9.604  103 1.229  105 4.145  102 4.580  104 3.968  104
Standard error 2.928  104 7.004  104 2.505  105 1.455  103 1.529  104 1.424  104

Fig. 9 e Observed shape changes with exposure time of two birds, i.e., a control group ðGC Þ and a treatment group ðGT Þ, (a and
f) Day 0, (b and g) Day 4, (c and h) Day 7, (d and i) Day 10, and (e and j) Day 13.

Finally, the proposed monitoring system not only provides variations of the extracted features variables were analysed in
early detection and prediction of disease occurrence but can terms of time after exposure to an infection and correlations
also automatically, continuously and non-intrusively take to the health status were developed. The results obtained
measurements throughout the lifespan of the birds. Thus, suggest that the proposed machine vision system can provide
reduced surveillance costs would benefit both the broiler and an early disease detection and prediction in broiler chicken.
poultry producer. However, this system needs to be validated on different types
of chicken breeds and infection types. Despite these limita-
tions, the proposed technique with acceptable results ach-
5. Conclusions ieved the objectives of this research study. It is of great
importance to detect an infection in a flock in a timely manner
This study proposes an early detection and prediction of a to allow stockmen and veterinarians to take necessary actions
disease occurrence in broiler chicken based on St , and gs . The to prevent huge loses and to keep a good animal welfare
 b i o s y s t e m s e n g i n e e r i n g 1 8 8 ( 2 0 1 9 ) 2 2 9 e2 4 2 241

status. Thus, this proposed system can be applied as a PLF Henchion, M., McCarthy, M., Resconi, V. C., & Troy, D. (2014). Meat
monitoring system for the entire life span of broilers. Addi- consumption: Trends and quality matters. Meat Science, 98(3),
tionally, this system can be integrated with other bio- 561e568. https://doi.org/10.1016/j.meatsci.2014.06.007.
Huang, J., Wang, W., & Zhang, T. (2019). Method for detecting
response monitoring systems such a lameness detection,
avian influenza disease of chickens based on sound analysis.
weight estimation, activity level, and behaviour detection Biosystems Engineering, 180, 16e24. https://doi.org/10.1016/
systems in a broiler chicken management system. j.biosystemseng.2019.01.015.
Jana, A. (2012). Kinect for windows SDK programming guide. Packt
Publishing Ltd.
Kleinbaum, D. G., & Klein, M. (2010). Maximum likelihood
Acknowledgments techniques: An overview. In Logistic regression (pp. 103e127).
Kongsro, J. (2014). Estimation of pig weight using a Microsoft
The project was funded by China National Key Research and Kinect prototype imaging system. Computers and Electronics in
Development project (Grant No. 2017YFE0114400). Agriculture, 109, 32e35. https://doi.org/10.1016/
j.compag.2014.08.008.
Kristensen, H. H., & Cornou, C. (2011). Automatic detection of
references deviations in activity levels in groups of broiler chickensea
pilot study. Biosystems Engineering, 109(4), 369e376. https://
doi.org/10.1016/j.biosystemseng.2011.05.002.
Kristensen, H. H., Prescott, N. B., Perry, G. C., Ladewig, J.,
Aydin, A. (2017a). Development of an early detection system for
Ersbøll, A. K., Overvad, K. C., et al. (2007). The behaviour of
lameness of broilers using computer vision. Computers and
broiler chickens in different light sources and illuminances.
Electronics in Agriculture, 136, 140e146. https://doi.org/10.1016/
Applied Animal Behaviour Science, 103(1e2), 75e89. https://
j.compag.2017.02.019.
doi.org/10.1016/j.applanim.2006.04.017.
Aydin, A. (2017b). Using 3D vision camera system to
Kurnianggoro, L., & Jo, K.-H. (2018). A survey of 2D shape
automatically assess the level of inactivity in broiler chickens.
representation: Methods, evaluations, and future research
Computers and Electronics in Agriculture, 135, 4e10. https://
directions. Neurocomputing, 300, 1e16. https://doi.org/10.1016/
doi.org/10.1016/j.compag.2017.01.024.
j.neucom.2018.02.093.
Bradshaw, R. H., Kirkden, R. D., & Broom, D. M. (2002). A review of
Manning, L., Chadd, S. A., & Baines, R. N. (2007). Key health and
the aetiology and pathology of leg weakness in broilers in
welfare indicators for broiler production. World’s Poultry
relation to welfare. Avian and Poultry Biology Reviews, 13(2),
Science Journal, 63(1), 46e62. https://doi.org/10.1017/
45e103. https://doi.org/10.3184/147020602783698421.
S0043933907001262.
Butcher, G. D., Jacob, J. P., & Mather, F. B. (1999). Common poultry
Mansournia, M. A., Geroldinger, A., Greenland, S., & Heinze, G.
diseases. PS47-Series of the veterinary medicine-large animal clinical
(2017). Separation in logistic regression: Causes,
Sciences department, UF/IFAS extension. Retrieved from: https://
consequences, and control. American Journal of Epidemiology,
edis.ifas.ufl.edu/pdffiles/PS/PS04400.pdf.
187(4), 864e870. https://doi.org/10.1093/aje/kwx299.
Caplen, G., Hothersall, B., Murrell, J. C., Nicol, C. J., Waterman-
Mast, J., Nanbru, C., Decaesstecker, M., Lambrecht, B.,
Pearson, A. E., Weeks, C. A., et al. (2012). Kinematic analysis
Couvreur, B., Meulemans, G., et al. (2006). Vaccination of
quantifies gait abnormalities associated with lameness in
chicken embryos with escape mutants of La Sota Newcastle
broiler chickens and identifies evolutionary gait differences.
disease virus induces a protective immune response. Vaccine,
PLoS One, 7(7), e40800. https://doi.org/10.1371/
24(11), 1756e1765. https://doi.org/10.1016/
journal.pone.0040800.
j.vaccine.2005.10.020.
Chansiripornchai, N., & Sasipreeyajan, J. (2006). Efficacy of live B1
Menesatti, P., Costa, C., Antonucci, F., Steri, R., Pallottino, F., &
or Ulster 2C Newcastle disease vaccines simultaneously
Catillo, G. (2014). A low-cost stereovision system to estimate
vaccinated with inactivated oil adjuvant vaccine for
size and weight of live sheep. Computers and Electronics in
protection of Newcastle disease virus in broiler chickens. Acta
Agriculture, 103, 33e38. https://doi.org/10.1016/
Veterinaria Scandinavica, 48(1), 2. https://doi.org/10.1186/1751-
j.compag.2014.01.018.
0147-48-2.
Mollah, M. B. R., Hasan, M. A., Salam, M. A., & Ali, M. A. (2010).
Damerow, G. (2016). The chicken health Handbook: A complete guide to
Digital image analysis to estimate the live weight of broiler.
maximizing flock health and dealing with disease. Storey
Computers and Electronics in Agriculture, 72(1), 48e52. https://
Publishing.
doi.org/10.1016/j.compag.2010.02.002.
Febrer, K., Jones, T. A., Donnelly, C. A., & Dawkins, M. S. (2006).
Møller, M. F. (1993). A scaled conjugate gradient algorithm for fast
Forced to crowd or choosing to cluster? Spatial distribution
supervised learning. Neural Networks, 6(4), 525e533. https://
indicates social attraction in broiler chickens. Animal
doi.org/10.1016/S0893-6080(05)80056-5.
Behaviour, 72(6), 1291e1300. https://doi.org/10.1016/
Mortensen, A. K., Lisouski, P., & Ahrendt, P. (2016). Weight
j.anbehav.2006.03.019.
prediction of broiler chickens using 3D computer vision.
Friedman, M. (1937). The use of ranks to avoid the assumption of
Computers and Electronics in Agriculture, 123, 319e326. https://
normality implicit in the analysis of variance. Journal of the
doi.org/10.1016/j.compag.2016.03.011.
American Statistical Association, 32(200), 675e701. https:// €a
€ s, I. A., de Paz, I. C. L. A., Baracho, M., dos, S., de
de Na
doi.org/10.1080/01621459.1937.10503522.
Menezes, A. G., de Lima, K. A. O., et al. (2010). Assessing
Friel, S., & Ford, L. (2015). Systems, food security and human
locomotion deficiency in broiler chicken. Scientia Agricola,
health. Food Security, 7(2), 437e451. https://doi.org/10.1007/
67(2), 129e135. https://doi.org/10.1590/S0103-
s12571-015-0433-1.
90162010000200001.
Glickman, M. E., Rao, S. R., & Schultz, M. R. (2014). False
OECD-FAO. (2017). Organisation for economic Co-operation and
discovery rate control is a recommended alternative to
development (OECD)/Food and agriculture organization of the
Bonferroni-type adjustments in health studies. Journal of
United Nations (FAO). 2017. Agricultural outlook 2017e2026:
Clinical Epidemiology, 67(8), 850e857. https://doi.org/10.1016/
Special Focus: Southeast Asia. Retrieved from: http://www.fao.
j.jclinepi.2014.03.012.
org/3/a-i7465e.pdf.
242 b i o s y s t e m s e n g i n e e r i n g 1 8 8 ( 2 0 1 9 ) 2 2 9 e2 4 2

Okinda, C., Lu, M., Nyalala, I., Li, J., & Shen, M. (2018). Asphyxia Tablante, N. L. (2013). Common poultry diseases and their prevention.
occurrence detection in sows during the farrowing phase by University of Maryland Extension. Retrieved from: https://
inter-birth interval evaluation. Computers and Electronics in extension.umd.edu/sites/extension.umd.edu/files/_docs/
Agriculture, 152, 221e232. https://doi.org/10.1016/ Common-Poultry-Diseases-and-Their-Prevention_Tablante_
j.compag.2018.07.007. 2013.pdf.
Otsu, N. (1979). A threshold selection method from gray-level Taheri, S. M., & Hesamian, G. (2013). A generalization of the
histograms. IEEE Transactions on Systems, Man, and Cybernetics, Wilcoxon signed-rank test and its applications. Statistical
9(1), 62e66. https://doi.org/10.1109/TSMC.1979.4310076. Papers, 54(2), 457e470. https://doi.org/10.1007/s00362-012-
Panagiotakis, C., & Argyros, A. (2016). Parameter-free modelling of 0443-4.
2D shapes with ellipses. Pattern Recognition, 53, 259e275. Thorp, B. H., & Duff, S. R. I. (1988). Effect of exercise on the
https://doi.org/10.1016/j.patcog.2015.11.004. vascular pattern in the bone extremities of broiler fowl.
Paul-Murphy, J. R., & Hawkins, M. (2014). Bird-specific Research in Veterinary Science, 45(1), 72e77. https://doi.org/
considerations: Recognizing pain behavior in pet birds. In 10.1016/S0034-5288(18)30897-X.
Handbook of veterinary pain management (3rd ed., pp. 536e554). Van der Stuyft, E., Schofield, C. P., Randall, J. M., Wambacq, P., &
Peiris, J. S. M., Cowling, B. J., Wu, J. T., Feng, L., Guan, Y., Yu, H., Goedseels, V. (1991). Development and application of
et al. (2016). Interventions to reduce zoonotic and pandemic computer vision systems for use in livestock production.
risks from avian influenza in Asia. The Lancet Infectious Computers and Electronics in Agriculture, 6(3), 243e265. https://
Diseases, 16(2), 252e258. https://doi.org/10.1016/S1473-3099(15) doi.org/10.1016/0168-1699(91)90006-U.
00502-2. Waterson, A. P., Pennington, T. H., & Allan, W. H. (1967). Virulence
Pereira, D. F., Miyamoto, B. C. B., Maia, G. D. N., Sales, G. T., in Newcastle disease virus: A preliminary study. British Medical
Magalha ~ es, M. M., & Gates, R. S. (2013). Machine vision to Bulletin, 23(2), 138e143. https://doi.org/10.1093/
identify broiler breeder behavior. Computers and Electronics in oxfordjournals.bmb.a070534.
Agriculture, 99, 194e199. https://doi.org/10.1016/ Welfare Quality®. (2009). Welfare Quality® assessment protocol
j.compag.2013.09.012. for poultry (broilers, laying hens). In Welfare Quality®
Perneger, T. V. (1998). What's wrong with Bonferroni adjustments. consortium, Lelystad, Netherlands (p. 114pp). Retrieved from:
BMJ, 316(7139), 1236e1238. https://doi.org/10.1136/ http://edepot.wur.nl/233471.
bmj.316.7139.1236. Wongsriworaphon, A., Arnonkijpanich, B., & Pathumnakul, S.
Rahman, S., Rabbani, M. G., Uddin, M. J., Chakrabartty, A., & (2015). An approach based on digital image analysis to
Her, M. (2012). Prevalence of avian influenza and Newcastle estimate the live weights of pigs in farm environments.
disease viruses using rapid antigen detection Kit in poultry in Computers and Electronics in Agriculture, 115, 26e33. https://
some areas of Bangladesh. Archives of Clinical Microbiology, 3(1). doi.org/10.1016/j.compag.2015.05.004.
https://doi.org/10.3823/248. Zhang, D., & Lu, G. (2004). Review of shape representation and
Rushton, J., Viscarra, R., Bleich, E. G., & McLeod, A. (2005). Impact description techniques. Pattern Recognition, 37(1), 1e19. https://
of avian influenza outbreaks in the poultry sectors of five doi.org/10.1016/j.patcog.2003.07.008.
South East Asian countries (Cambodia, Indonesia, Lao PDR, Zhuang, X., Bi, M., Guo, J., Wu, S., & Zhang, T. (2018). Development
Thailand, Viet Nam) outbreak costs, responses and potential of an early warning algorithm to detect sick broilers.
long term control. World’s Poultry Science Journal, 61(3), Computers and Electronics in Agriculture, 144, 102e113. https://
491e514. https://doi.org/10.1079/WPS200570. doi.org/10.1016/j.compag.2017.11.032.
Sims, L. D. (2008). Risks associated with poultry production Zhuang, X., & Zhang, T. (2019). Detection of sick broilers by digital
systems. International Conference Poultry in the Twenty-First image processing and deep learning. Biosystems Engineering,
Century, 1, 24. Retrieved from: http://www.fao.org/WaICENT/ 179, 106e116. https://doi.org/10.1016/
FAOINFO/AGRICULT/againfo/home/events/bangkok2007/ j.biosystemseng.2019.01.003.
docs/part2/2_1.pdf.