Food Bioprocess Technol (2013) 6:2503–2513
DOI 10.1007/s11947-012-0933-3
ORIGINAL PAPER
Detecting Fertility and Early Embryo Development of Chicken
Eggs Using Near-Infrared Hyperspectral Imaging
L. Liu & M. O. Ngadi
Received: 15 February 2012 / Accepted: 9 July 2012 / Published online: 2 August 2012
# Springer Science+Business Media, LLC 2012
Abstract Early detection of infertile and non-hatchable
eggs would benefit hatcheries and poultry breeding farms
by saving space, handling costs, and preventing contamina-
tion from exploder eggs. Therefore, it would be advanta-
geous to the hatchery industry of developing a non-
destructive, rapid, and accurate method to detect the fertility
and embryo development of eggs. For this purpose, a near-
infrared hyperspectral imaging system was developed to
detect fertility and early embryo development. A total of
174 white-shell chicken eggs including 156 fertile eggs and
18 infertile eggs were used in this study and all eggs were
incubated in a commercial incubator for 4 days. Hyperspec-
tral images were captured for all eggs on each day of
incubation. After imaging on each day, developing embryos
in randomly selected eggs were stopped by injecting sodium
azide (NaN3). All the eggs were divided into two classes,
fertile eggs and non-fertile eggs (including infertile eggs and
dead embryos), and the data set of each class varied with
day of incubation. The region of interest (ROI) of each
hyperspectral image was segmented and the image texture
information was extracted from the ROI of spectral images
using Gabor filters. Two types of spectral transmission
characteristics termed MS and MG, were obtained by aver-
aging the spectral information of ROI and Gabor-filtered
ROI, respectively. The dimensionality of the spectral trans-
mission characteristics were reduced by PCA. The first three
PCs were used for K-means clustering, as well as the first
three bands with maximum responses of each spectral trans-
mission characteristic. The best classification results were
100 % at day 0, 78.8 % at day 1, 74.1 % at day 2, 81.8 % at
L. Liu : M. O. Ngadi (*)
Department of Bioresource Engineering, McGill University,
Macdonald Campus 21, 111 Lakeshore Road,
Ste-Anne-de-Bellevue, QC, Canada H9X 3V9
e-mail: michael.ngadi@mcgill.ca
day 3, and 84.1 % at day 4. A perfect detection of fertility
prior to incubation was obtained using only the first three
bands of maximum responses of MS. The classification
results suggested the usefulness of the image texture infor-
mation for detection of early embryo development. Promis-
ing results were also obtained when only the first three
bands with maximum response of spectral transmission
characteristics were used, which indicated the potential in
applying hyperspectral imaging techniques to develop a
real-time system for detecting fertility and early embryo
development of chicken eggs.
Keywords Hyperspectral imaging . Egg . Fertility . Embryo
development . Image texture . Gabor filter . PCA . K-means
clustering
Introduction
Hatchability of eggs is a critical economic factor for hatch-
eries and poultry breeding farms. It is influenced by many
factors such as breeder flock problem, fertility, egg han-
dling, etc. Early detection of non-fertile and non-hatchable
eggs would allow hatcheries to remove them before transfer
to the incubator/hatcher, thus saving space, handling costs,
and contamination from exploder eggs. In the poultry in-
dustry, candling is widely used to assess flock fertility so
that infertile eggs and dead embryos can be removed from
the incubator. However, since candling is labor consuming,
only few eggs in the incubator are candled for determination
of flock fertility, which means most non-fertile and non-
hatchable eggs will remain in the incubator until they are
moved to the hatcher. Therefore, the development of an
efficient, non-destructive, and accurate method for detecting
the fertility and the embryo development of eggs would be
advantageous to the hatchery industry.
2504
Some researchers have attempted to detect fertility and
embryo development with computer vision technology. Das
and Evans (1992a, b) developed machine vision systems to
determine fertility of hatching eggs by combining image
histogram characterization with neural network classifiers.
The classification results reached 93.9 % at day 4 of incu-
bation, 93.5 % at day 3, and 67.6 % at day 2. Bamelis et al.
(2002) applied the spectrophotometric method to detect
embryo development of chicken eggs during the first 12 days
of incubation. Embryo development of chicken eggs was
stopped by injecting sodium azide at different times during
incubation. They concluded that the embryo development
can be detected using visible light transmission from
108 h (day 4.5) of incubation, which is not directly
linked with the formation of blood, but with the forma-
tion of sub-embryonic fluid which causes the yolk to
become translucent.
Recently, the advanced hyperspectral imaging technique
has been applied to detect egg fertility and embryo devel-
opment (Lawrence et al. 2006; Smith et al. 2008). Hyper-
spectral imaging combines the conventional 2-D digital
imagery with spectroscopy. A hyperspectral image contains
not only spatial information, but also spectral information
for each pixel in the image. This information is extracted to
form a 3-D “hypercube” which can be analyzed to ascertain
minor and/or subtle physical and chemical features in an
object. Thus, a hyperspectral image can be used not only to
detect physical and geometric characteristics such as color,
size, shape, and texture but also to extract some intrinsic
chemical information such as water, fat, and protein. The
technology clearly has a promising future for food quality
and safety inspections (Polder et al. 2002; Lee et al. 2005).
In Lawrence et al. (2006), the hyperspectral imaging
technique was used for detecting early embryo development
of hatching eggs at the first 3 days of incubation. Both
Mahalanobis distance (MD) classification and partial least
squares regression (PLSR) were used to build the prediction
model based on the hyperspectral images of hatching eggs.
The authors claimed that the embryo classification results
based on the MD model with all eggs over all the incubation
days were about 96 % at day 0, 92 % at day 1, and 100 % at
days 2 and 3, while the PLSR model based on eggs only
from days 0 and 3 produced 100 % classification accuracy at
days 2 and 91.7 % at day 1. The extremely promising results
encouraged more research in this area. The same research
group later introduced infertile eggs to evaluate the hyper-
spectral imaging system and a predictive modeling tech-
nique for determining fertility prior to incubation and
embryo development during the first 3 days of incubation
(Smith et al. 2008). However, the prediction model pro-
duced much lower classification results both on validation
data (71 % for day 0, 63 % for day 1, 65 % for day 2, and
83 % for day 3) and verification data (50.8 % overall). The
Food Bioprocess Technol (2013) 6:2503–2513
results in both studies indicated that development of proper
HSI-based techniques for accurate prediction is necessary
for detecting early embryo development and/or fertility prior
to incubation.
When observing the contents of the egg in a dark
room with candling, the living embryo appears as a dark
spot in the large end of the egg surrounded by a faint
outline of blood vessels, while dead embryos sometimes
appear as a ring or smear of blood in the egg or a dark
spot dried to the inside of the shell. The infertile egg
brightly transmits light in comparison. The appearance of
blood vessels in embryo indicated that spectral imaging
in the near-infrared (NIR) region might produce better
detection results than in the visible light. In addition,
most organic materials have excellent transmittance and
reflectance properties within the NIR range (Williams
2003). Thus, the possibility of using image texture fea-
ture for detecting fertility and embryo development
should be considered as they could possibly bring im-
provement for the fertility and embryo detection.
A 2-D Gabor filter is a linear filter which is a sinu-
soidal modulated Gaussian function in the spatial do-
main. The 2-D Gabor filter is well-known to have the
ability to achieve certain optimal joint localization prop-
erties in the spatial domain and in the spatial frequency
domain (Daugman 1985). The kernel of 2-D Gabor filter
is similar to the response of the 2-D receptive field
profiles of simple cells in the visual cortex of mamma-
lian brains (Daugman 1980) and thus is similar to those
of the human visual system. This ability exhibits desir-
able characteristics of capturing salient visual properties
such as spatial localization, orientation selectivity, and
spatial frequency. These characteristics make it very ef-
fective and particularly appropriate for image texture
representation, analysis, and discrimination (Manjunath
and Ma 1996; Clausi and Ed Jernigan 2000).
The objective of this study was to develop an efficient
and accurate system based on hyperspectral imaging tech-
nique to determine fertility of hatching eggs prior to incu-
bation and embryo development during the first 4 days of
incubation.
Materials and Methods
Eggs and Sample Evolution
A total of 174 freshly laid white eggs including 156 fertile
eggs and 18 infertile eggs were collected from the McGill
University Farm in five batches over 3 months. On day 0
(just prior to incubation), all eggs were first imaged by the
hyperspectral imaging system and then a total of 20 fertile
eggs over five batches were randomly selected for inducing
Food Bioprocess Technol (2013) 6:2503–2513 2505

dead embryo (stopping embryo development). Each of the

20 fertile eggs was injected with a 100 μl of a 5 % sodium
azide (NaN3) solution using a sterile needle through a tiny
hole drilled at the sharp end of the egg (Bamelis et al. 2002).
After imaging and injecting, all eggs were immediately
incubated in a 2362N Circulated Air Hova-Bator incubator
(G.Q.F. Manufacturing Co., Savannah, GA, USA) at 100 °F
and 55 % relative humidity, and were turned every hour. On
days 1, 2, 3, and 4 of incubation, eggs were removed for
imaging in sequence and then immediately replaced in the
incubator with a process of no more than 3 min. After all
eggs were imaged on each incubation day, a total of 20
fertile eggs were randomly selected, removed, and injected
with NaN3, and then replaced in the incubator. After 7 days
of incubation, eggs were candled and broken out to deter-
mined fertility and embryo viability. Out of the total number
of eggs studied, one, two, and one eggs that were injected on
days 1, 2, and 3, respectively, were broken.
Table 1 lists the sample evolution during the first 4 days
of incubation. A total of six groups were obtained on day 4
of incubation: fertile eggs (n076), infertile eggs (n018), Fig. 1 The hyperspectral imaging system used in the study
eggs injected (i.e., stopped embryo development) on days 0
(n020), 1 (n019), 2 (n018), and 3 (n019). Note that the
number of eggs with stopped embryo development varied Headwall Photonics Inc. USA), an InGaAs camera
because of broken eggs. For each day of incubation, there connected to the spectrograph, a conveyer (Donner
were two classes: fertile eggs and non-fertile eggs. In the 2200 series, Donner Mfg. Corp., USA) driven by a
rest of the paper, non-fertile eggs specifically include infer- stepping motor (MDIP22314, Intelligent motion system
tile eggs and dead embryos. The numbers of eggs in the two Inc., USA) with a user-defined speed, a tungsten halo-
classes varied with the time of incubation as shown in gen lamp (50 W) providing back illumination to eggs,
Table 1. an enclosure supporting the system, a data acquisition
and preprocessing software (Hyperspec, Headwall Pho-
Hyperspectral Imaging System and Image Preprocessing tonics Inc. USA) and a PC.
The system collected spectral images in a wavelength
Figure 1 shows the laboratory near-infrared (NIR) hyper- range between 900 and 1,700 nm with a spectral resolution
spectral imaging system used in the study. This imaging of 2.8 nm. Egg samples vertically placed on a wood board
system consisted of a line-scan spectrograph (Hyperspec™, were conveyed to the field of view of the camera with a
predefined speed which was decided by trial-to-error to
avoid image distortion. The halogen lamp was placed right
Table 1 Egg sample evolution during the first 4 days of incubation underneath the camera. The two parameters of the prepro-
cessing software, i.e., system integration time and the
Incubation time Before imaging After imaging frames used for calculating the output signal, were set at
5,000 ms and eight, respectively. The output hypercube is
Total Fertile Non-fertile eggs Injected eggs
eggs eggs (infertile eggs, (stopped embryo, 800 rows×320 columns×167 bands.
dead embryo) broken eggs ) Image preprocessing was conducted to calibrate the spec-
tral images by percent transmission. A dark and a white
Day 0 174 156 18 (18, 0) 20 (20, 0) image was obtained for calculating the percent transmission
Day 1 174 136 38 (18, 20) 20 (19, 1) of the output spectral images by covering the lens with a
Day 2 173 116 57 (18, 39) 20 (18, 2) cap, and allowing the light to go through the lens, respec-
Day 3 171 96 75 (18, 57) 20 (19, 1) tively. Thus, the pixel values of each plane of the output
Day 4 170 76 94 (18, 76) – hypercube are between 0 (corresponding to the dark image
There are two classes (fertile vs. non-fertile) on each day of incubation
that means no light transmitted) and 1 (corresponding to the
based on fertility and six groups for all egg samples according to the white image that means all light from the halogen lamp
injection time (marked in italics) transmitted).
2506 Food Bioprocess Technol (2013) 6:2503–2513

Fig. 2 An example of spectral

images at the same wavelength
(1,076 nm) on different days of
incubation

Day 0 Day 1 Day 2 Day 3 Day 4

Gabor Filter mask for each egg and then to segment each egg from the
original image.
The objective of using Gabor filter was to extract useful In order to extract useful and helpful image texture
image texture information to facilitate the classification information of eggs, Gabor filters were applied to the
of fertile and non-fertile eggs (infertile eggs and dead ROI of hyperspectal images. An oriented Gabor filter
embryos). All operations in this section were performed G0, which is a Gaussian function modulated by an
using MATLAB 7.3.0 (The MathWorks, Inc., MA, oriented harmonic function, can be defined as follows
USA). (Ma et al. 2002):
A hyperspectral image captured by the imaging system
2 
1 x þ y2
described in “Hyperspectral Imaging System and Image G0 ðx; y; f ; σ; θÞ ¼ exp 
2pσ2 2σ2
Preprocessing” normally consists of four eggs. The selection
of region of interest (ROI) of a sample image is to create a  cos½2pf ðx cos θ þ y sin θÞ; ð1Þ

Fig. 3 Selection of ROI. a The

spectral image at 1,076 nm, b
the mask for ROI, and c four
separated regions of interest

(a) (b) (c)

Food Bioprocess Technol (2013) 6:2503–2513 2507

Fig. 4 The region of interest

and the corresponding Gabor-
filtered images

Filtered ROI

Circular Gabor
Filter

ROI

Oriented Gabor
Filter

Filtered
ROI

0° 45° 90° 135°

where (x, y) is the coordinate of point in 2-D space, f is In order to make the Gabor filter more robust against
the frequency of the sinusoidal wave, σ is the standard differences in brightness, both Gabor filters were tuned to
deviation of the Gaussian envelope, and θ controls the zero DC (direct current, i.e., the baseline) with the application
orientation of the Gabor filter. When only the spatial of the following formula (Zhang et al. 2003):
frequency information accounts for the major differences
among texture, an isotropic Gabor filter G1 can be used P
n P
n
G½i; j
with the following form: e
Gg ¼ Gg 
i¼n j¼n
; ð3Þ

2  ð2n þ 1Þ2
1 x þ y2
G1 ðx; y; f ; σ; θÞ ¼ exp 
2pσ2 2σ2 where g00,1, (2n+1)2 is the size of the filter. The adjusted
h pffiffiffiffiffiffiffiffiffiffiffiffiffiffii e was used to convolute the ROI of each sample
Gabor filter G
 cos 2pf x2 þ y2 ð2Þ
image, respectively.

Fig. 5 Typical spectral images

(at the wavelength of 1,076 nm)
of different groups at day 4. a
Infertile eggs, dead embryos at
b day 0, c day 1, d day 2, e
day 3, and f fertile eggs
2508 Food Bioprocess Technol (2013) 6:2503–2513

Spectral Transmission Characteristics
Two types of spectral transmission characteristics MS and MG
were used for further data analysis. MS was the mean spectra
of ROI (i.e., mean value of all pixels in ROI in the current
wavelength) over the spectral range of 900–1,700 nm, while
MG was the mean spectra of the Gabor-filtered ROI over the
same spectral range. Each spectral transmission characteristic
was a 167×1 vector and was denoised by a 5×1 mean filter. A
baseline filtering method was also used to remove baseline
from the original spectral transmission characteristics. In the
rest of the paper, the spectral transmission characteristics MS
and MG refer to the characteristics without noise and baseline.
Data Analysis
In the stage of data classification, spectral transmission char-
acteristics obtained in “Hyperspectral Imaging System and

Fig. 6 Typical curves of

spectral transmission
characteristics a MS, b MG1,
and c–f MG00–03 for different
groups of eggs at day 4
Food Bioprocess Technol (2013) 6:2503–2513
Image Preprocessing” were first projected onto a lower di-
mensional linear space to extract useful features and then the
extracted features were classified by using typical unsuper-
vised classification methods, i.e., K-means clustering. All
operations in this section were performed using MATLAB
7.3.0 (The MathWorks, Inc., Mass., USA).
Principal component analysis (PCA), also known as Kar-
hunen–Loeve transform, is widely used for dimensionality
reduction, loss data compression, and feature extraction.
This technique projects the data orthogonally onto a lower
Fig. 7 Evolution of the peak
transmission of a MS, b MG1,
and c–f MG00–03 in different
groups. Mean of each group at
different time are presented,
error bars show SEM
2509
dimensional linear space such that the variance of the pro-
jected data is maximized. In this study, PCA was employed
to extract no redundant features from spectral information
for efficient and effective classification.
The K-means clustering (Hartigan and Wong 1979) algo-
rithm is a well-known unsupervised classification method
which generates a specific number of disjoint, flat (non-hierar-
chical) clusters by minimizing the sum, over all clusters, of the
within-class distances from each object to its cluster center. The
K-means method has been widely used in image analysis and
2510
computer vision due to its simplicity and speed. In this study,
the egg samples were classified using K-means clustering.
Results and Discussions
Spectral Transmission Characteristics
Figure 2 shows the evolution of spectral images of the same
set of eggs captured by the hyperspectral imaging system,
where all eggs were fertile and all spectral images at the
same wavelength of 1,076 nm. The region of interest (ROI)
of each hyperspectral image (Fig. 3a), i.e., the egg areas,
was segmented using a predefined mask (Fig. 3b. A pixel in
Fig. 3a remains unchanged if the corresponding pixel in
Fig. 3b is marked white, otherwise set to zero.) and four
regions of interest (Fig. 3c) were obtained for each hyper-
spectral image, as shown in Fig. 3. Each ROI was filtered by
an isotropic Gabor filter G e 1 and four oriented Gabor filters,
e e e e
i.e., G00 , G01 , G02 , G03 , along 0,π/4,π/2,3π/4, respectively,
and the corresponding Gabor-filtered ROI images were
obtained as illustrated in Fig. 4, where f00.05, σ020, and
n041 for all Gabor filters based on trial and error.
Typical spectral images of different groups at day 4 are
shown in Fig. 5 and the corresponding typical curves of
spectral transmission characteristics, MS and MG, are
shown in Fig. 6. All spectral transmission characteristics
are very close to 0 in the range of 900–950 and 1,155–
1,700 nm and thereby no deviation can be found among
different groups. The largest deviation between the 6 groups
examined in this study can be found at the wavelength of
1,076 nm where the peak value for every spectral transmis-
sion characteristics were produced, no matter MS or MG
were used. The peak observed in the NIR region was due to
the water and fat absorption bands related to O–H and C–H
stretching second where the NIR light has the high penetra-
tion power through water and fat (Anderson et al. 2006).
The peak transmission value for fertile eggs at day 4 was
minimal among all groups, which indicates that the values
of spectral transmission characteristics around the peak
could be used to discriminate the fertile eggs from non-
fertile eggs. In addition, the deviation between fertile eggs
and injected embryos at day 3 (injection at day 3, the last
injected embryos in this study) was larger in MG than MS.
This indicated the possibility of improvement of detecting
recent dead embryos (within 24 h) using image texture
information.
Evolution of Spectral Transmission
The evolution of average peak values of spectral transmis-
sion characteristics for different groups is shown in Fig. 7,
where the marker face presents the average value and error
Food Bioprocess Technol (2013) 6:2503–2513
bars illustrate the standard error of mean (SEM). For all
spectral transmission characteristics of infertile eggs, a sharp
decrease of the average peak transmission happened at day 1
of incubation versus prior to incubation, while the average
peak transmission was constant (MS, MG00, and MG01) or
change little (MG02, MG03, and MG1) after 24 h of incu-
bation. In contrast, the average peak transmission of fertile
eggs varies with different types of spectral transmission
characteristics. The average peak value of MS for fertile
eggs kept decreasing during incubation and dropped down
more at days 3 and 4 than days 1 and 2, while the average
peak transmission of MG00–03 kept increasing during the
first 72 h of incubation and then significantly decreased at
day 4 that is 1 day earlier than the observation of a signif-
icant decrease of the 577:610 nm ratio in fertile eggs
reported in Bamelis et al. (2002). A big difference of the
average peak transmission between fertile eggs and infertile
eggs was observed at day 0 (prior to incubation) and day 4
(after 96 h of incubation), which indicated that fertile and
infertile eggs could be totally classified at the 2 days.
The groups of treatments after different periods of incu-
bation show various evolution profiles with different injec-
tion time. The developing embryos stopped at days 0 and 1
produced similar evolution profiles with the infertile eggs
except no significant decrease at day 1, while the eggs
injected at day 3 had a similar profile as fertile eggs except
that signification decrease at day 4 was not so sharp as
fertile eggs. The embryos injected at day 2 showed a mixed
profile which had similar evolution as fertile eggs during the
first 3 days and no significant decrease happened at day 4.
For all spectral transmission characteristics, the later
injected eggs produced the larger difference with the first
injected group, which is same as the result in Bamelis et al.
Fig. 8 Cumulative percent variance explained by principal compo-
nents of different spectral transmission characteristics over all days
Food Bioprocess Technol (2013) 6:2503–2513 2511

Fig. 9 Distribution of egg

samples in the space spanned
by the first three PCs which
were obtained from MS and
MG1 at days 0 and 4 of
incubation. Only fertile (with
red marker) and infertile (with
blue marker) eggs are shown on
day 0, while fertile eggs, dead
embryos (with green marker),
and infertile eggs are shown on
day 4. The view of each plot
was adjusted for the possible
best discrimination of classes

(a) MS on Day 0 (b) MS on Day 4

(c) MG1 on Day 0 (d) MG1 on Day 4

(2002). The average peak transmission became constant
after 24 h of injection.
PCA
The cumulative percentage of variance explained by princi-
pal components of MBRS and MBRG over all incubation
days is shown in Fig. 8. The first three PCs explained at
least 99 % of the variance of data for all spectral transmis-
sion characteristics over all days and thereby were selected
for classification. Figure 9 shows the principal component
score plots of the first three PCs for MS and MG1 (along the
vertical direction) at days 0 and 4 of incubation (along the
horizontal direction) as well as the percentage of variance
explained by each PC. A sample score is the map of the
original data into the space spanned by the principal com-
ponents. Principal component score plots visualize the dis-
crimination of classes based on the plotted PC scores. Three
different data sets, i.e., dead embryos, fertile eggs, and
infertile eggs, are shown in Fig. 9. The data sets of fertile

Table 2 Accuracy (percent) of K-means clustering based on the first three PCs

Incubation MS MG00 MG01 MG02 MG03 MG1

time
Fertile Non- Total Fertile Non- Total Fertile Non- Total Fertile Non- Total Fertile Non- Total Fertile Non- Total
fertile fertile fertile fertile fertile fertile

Day 0 76.3 100 78.8 75.7 100 78.2 75.7 100 78.2 62.5 100 66.5 75.0 100 77.7 75.7 100 78.2
Day 1 63.6 68.4 64.7 63.6 71.1 65.3 64.4 63.2 64.1 66.7 50.0 62.9 63.6 65.8 64.1 65.2 60.5 64.1
Day 2 77.9 45.6 67.1 97.4 24.6 72.9 99.1 24.6 74.1 82.3 15.8 60.0 100 21.1 73.5 99.1 24.6 74.1
Day 3 77.9 68.0 73.5 72.6 54.7 64.7 72.6 54.7 64.7 72.6 54.7 64.7 72.6 54.7 64.7 72.6 53.3 64.1
Day 4 93.4 66.0 78.2 89.5 75.5 81.8 89.5 71.3 79.4 86.8 75.5 80.6 86.8 73.4 79.4 90.8 78.7 84.1

The overall best classification results for all eggs on the corresponding incubation day are in italics
2512 Food Bioprocess Technol (2013) 6:2503–2513

Table 3 Accuracy (percent) of K-means clustering based on the first three bands of maximum responses

Incubation MS MG00 MG01 MG02 MG03 MG1

time
Fertile Non- Total Fertile Non- Total Fertile Non- Total Fertile Non- Total Fertile Non- Total Fertile Non- Total
fertile fertile fertile fertile fertile fertile

Day 0 100 100 100 100 77.8 97.7 78.3 88.9 79.4 93.4 11.1 84.7 73.7 94.4 75.9 98.7 100 98.8
Day 1 63.6 68.4 64.7 63.6 71.1 65.3 64.4 68.4 65.3 74.2 63.2 71.8 60.6 81.6 65.3 77.3 65.8 74.7
Day 2 59.3 50.9 56.5 51.3 63.2 55.3 63.7 61.4 62.9 81.4 36.8 66.5 69.0 47.4 61.8 62.8 54.4 60.0
Day 3 77.9 69.3 74.1 64.2 64.0 64.1 83.2 56.0 71.2 66.3 40.0 54.7 84.2 49.3 68.8 83.2 48.0 67.7
Day 4 92.1 69.2 79.4 88.2 77.7 82.4 88.2 72.3 79.4 86.8 76.6 81.2 86.8 74.5 80.0 90.8 76.6 82.9

The overall best classification results for all eggs on the corresponding incubation day is in italics

eggs and dead embryos vary with different days of incuba-
tion. Non-fertile eggs consist of only infertile eggs on day 0
(prior to incubation) and include infertile eggs and dead
embryos on the other days of incubation. The size of each
data set on different days of incubation is listed in Table 1.
K-Means Clustering
An unsupervised classification method, K-means clustering,
was employed to investigate the ability of spectral transmission
characteristics to discriminate fertile and non-fertile eggs. The
first three PCs of each spectral transmission characteristics
were used for classification and the corresponding results were
shown in Table 2. Since the largest deviation among
different groups was found around the waveband where
the maximal responses of MS and MG were produced
(as shown in Fig. 6), the first three bands with maxi-
mum responses (i.e., 1,066, 1,071, and 1,076 nm) of
each spectral transmission characteristic were also used
for classification and the K-means clustering results
were listed in Table 3.
A very promising result, i.e., 100 % accuracy for classi-
fication of fertile and infertile eggs prior to incubation
(day 0), was obtained when using the first three bands of
maximum responses of MS. During the incubation days
(day 1–4), the highest classification results over different
spectral transmission characteristics between Tables 2 and 3
were 74.7 % at day 1 (by first three bands of MG1), 74.1 %
at day 2 (by first three PCs of MG1), 74.1 % at day 3 (by
first three bands of MS), and 84.1 % at day 4 (by first three
PCs of MG1). The results showed that MS and MG1 pro-
duced better results than other spectral transmission charac-
teristics no matter the first three PCs or first three bands of
maximum responses used, which indicated the possibility of
obtaining better classification results when combining MS
and MG1.
Table 4 shows the K-means clustering results using the
first three bands of maximum responses of MS and MG1 (a
total of six bands). Higher results were produced on day 1
(78.8 %) and day 3 (81.8 %) compared to the first three
bands with maximum response of either MS or MG1 used,
while a slightly lower result on day 0 (99.4 %). Comparable
results on day 2 (70.0 %) and day 4 (81.2 %) were obtained
with the combination of first three bands of MS and MG1,
which were a little lower than the results from the first three
PCs of MG1 (74.1 % at day 2 and 84.1 % at day 4). One
possible explanation is that the first three PCs which are
obtained by all of 167 wavebands include more information
than the first three bands. However, the use of first three
bands with maximum response is more suitable for a real-
time system and thereby for the industrial application.
Comparing the results listed in Tables 4 and 3, the combi-
nation of MS and MG1 produces much better results at day 1
(78.8 vs. 74.7 %), day 2 (70.0 vs. 66.5 %), and day 3 (81.8 vs.
74.1 %) than the use of either MS or MG1, while the classifi-
cation accuracy obtained by this combination is slightly lower
than MS at day 0 (99.4 vs. 100 %) and MG1 at day 4 (81.2 vs.
82.9 %). This suggested that the image texture information is
more suitable for detecting early embryo development than
detection of fertility especially prior to incubation. The lower
classification results at days 1 and 2 indicated that the embryo
development is more difficult to be detected during the first
2 days of incubation, which is consistent with the fact that
Table 4 Accuracy (percent) of K-means clustering based on the first
three bands with maximum responses of MS and MG1
Incubation time First three maximal response bands of MS+MG1
Fertile Non-fertile Total
Day 0 99.3 100 99.4
Day 1 83.3 63.2 78.8
Day 2 86.7 36.8 70.0
Day 3 89.5 72.0 81.8
Day 4 90.8 73.4 81.2
The overall best classification results for all eggs on the corresponding
incubation day are in italics
Food Bioprocess Technol (2013) 6:2503–2513
blood formation starts in the developing embryo from day 2
(Romanoff 1960; Bodemer 1970).
Conclusions
The evolution of average peak values of spectral transmis-
sion characteristics showed that the later injected eggs pro-
duced the larger difference with the first injected group and
the average peak transmission of MS and MG remained
constant or change little after 24 h of injection. The best
unsupervised classification results for all eggs were 100 %
at day 0, 78.8 % at day 1, 74.1 % at day 2, 81.8 % at day 3,
and 84.1 % at day 4. The 100 % classification result at day 0
means a perfect detection of fertility prior to incubation. The
lower classification results at days 1 and 2 indicates that the
detection of embryo development is more difficult during
the first 2 days of incubation, which is consistent with the
period of blood formation in the developing embryo. The
classification results show that the image texture informa-
tion is more useful to detect early embryo development than
to detect fertility of chicken eggs especially prior to incuba-
tion. In addition, the promising results produced by the first
three bands with maximum response (i.e., 1,066, 1,071, and
1,076 nm) of MS (spectral transmission characteristic) and
MG1 (isotropic Gabor filter-based spectral transmission
characteristic) indicates the potential of hyperspectral imag-
ing techniques in building a real-time system for detection
of fertility and early embryo development of chicken eggs.
References
Anderson, R. R., Farinelli, W., Laubach, H., Manstein, D., Yaroslavsky,
A. N., Gubeli, J., et al. (2006). Selective photothermolysis of lipid-
rich tissues: a free electron laser study. Lasers in Surgery and
Medicine, 38, 913–919.
Bamelis, F. R., Tona, K., De Baerdemaeker, J. G., & Decuypere, E. M.
(2002). Detection of early embryonic development in chicken
2513
eggs using visible light transmission. British Poultry Science.,
43, 204–212.
Bodemer, C. W. (1970). Modern embryology. London: Holt, Rinehart
& Winston.
Clausi, D. A., & Ed Jernigan, M. (2000). Designing Gabor filters for
optimal texture separability. Pattern Recognition., 33(1), 1835–
1849.
Das, K., & Evans, M. D. (1992a). Detecting fertility of hatching eggs
using machine vision: I. Histogram characterization method.
Transactions of ASAE, 35, 1335–1341.
Das, K., & Evans, M. D. (1992b). Detecting fertility of hatching eggs
using machine vision: II. Neural network classifiers. Transactions
of ASAE, 35, 2035–2041.
Daugman, J. G. (1980). Two-dimensional spectral analysis of cortical
receptive field profiles. Vision Research., 20, 847–856.
Daugman, J. G. (1985). Uncertainty relation for resolution in space,
spatial frequency, and orientation optimized by two-dimensional
visual cortical filters. Journal of the Optical Society of America A,
2(7), 1160–1169.
Hartigan, J. A., & Wong, M. A. (1979). A k-means clustering algo-
rithm. Applied Statistics, 28(1), 100–108.
Lawrence, K. C., Smith, D. P., Windham, W. R., Heitschmidt, G. W., &
Park, B. (2006). Egg embryo development detection with hyper-
spectral imaging. Internation Journal of Poultry Science, 5(10),
964–969.
Lee, K. J., Kang, S., Kim, M. S., & Noh, S. H. (2005). Hyperspectral
imaging for detecting defect on apples. ASAE Paper No. 053075,
St Joseph, Michigan.
Ma, L., Tan, T., Wang, Y., & Zhang, D. (2002). Personal identification
based on iris texture analysis. IEEE Transactions on Pattern
Recognition and Machine Intelligence, 25(12), 1519–1533.
Manjunath, B. S., & Ma, W. Y. (1996). Texture feature for browing and
retrieval of image data. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 18(8), 837–842.
Polder, G., Heijden, G. W. A. M., & Young, I. T. (2002). Spectral
image analysis for measuring ripeness of tomatoes. Transactions
of ASAE, 45(4), 1155–1161.
Romanoff, A. L. (1960). The avian embryo. New York: Macmillan.
Smith, D. P., Lawrence, K. C., & Heitschmidt, G. W. (2008). Fertility
and embryo development of broiler hatching eggs evaluated with
a hyperspectral imaging and predictive modeling system. Inter-
national Journal of Poultry Science., 7(10), 1001–1004.
Williams, P. (2003). Near-infrared technology: getting the best out of
light. 2.0.2 ed. PDK Projects, Inc., Nanaimo, British Columbia,
Canada.
Zhang, D., Kong, W. K., You, J., & Wong, M. (2003). Online palm-
print identification. IEEE Transactions on Pattern Recognition
and Machine Intelligence, 25(9), 1041–1050.