Semi-automated Identification of Leopard Frogs
∗
Dijana Petrovska-Delacr´etaz
†,1
, Aaron Edwards
2
, John Chiasson
2
, G´erard Chollet
2,3
and David S. Pilliod
4,5
1
Electronics and Physics (EPH), Department of the Mines Telecom SudParis, CNRS Samovar, Paris, France
2
ECE Dept, Boise State University, Boise, 83725, ID, U.S.A.
3
LTCI of CNRS, Institut Mines-T´el´ecom, Paris, France
4
U.S. Geological Survey, Forest and Rangeland Ecosystem, Science Center Boise, Boise, 83706, Idaho, U.S.A.
5
Graduate Faculty of the Department of Biological Sciences at Boise State University, Boise, 83725, ID, U.S.A.
Keywords:
Animal Biometrics, Automatic Identification, Frogs, Principal Component Analysis.
Abstract:
Principal component analysis is used to implement a semi-automatic recognition system to identify recaptured
northern leopard frogs (Lithobates pipiens). Results of both open set and closed set experiments are given.
The presented algorithm is shown to provide accurate identification of 209 individual leopard frogs from a
total set of 1386 images.
1 INTRODUCTION
Identification of individual frogs in wild populations
is important for biologists who are conducting de-
mography studies used to evaluate the status and
trends of endangered species. Wildlife biologists have
used various methods to identify individuals in the
wild, most of which involve some type of permanent
or temporary mark or tag. These identification meth-
ods, while often reliable, may pose health risks to an-
imals and thus there is a need for non-harmful alter-
natives. One of the most intriguing alternatives for
animal identification is photography.
Photographically-based frog identification is con-
ducted in the following manner. Biologists capture
wild frogs from a study site (e.g., a pond), photograph
them, and then release them back into the population.
Later (e.g., days, weeks, months, or even annually),
biologists return to the study site and capture another
group of frogs, photograph them, and return them to
the population. The biologists then try to match in-
dividual frogs from the second group (set) to indi-
viduals caught during the previous visit (or all pre-
∗
Any use of trade, product, or firm names is for descrip-
tive purposes only and does not imply endorsement by the
U.S. Government.
†
This work was done during a sabbatical stay of Dijana
Petrovska-Delacr´etaz in the ECE Dept at Boise State Uni-
versity, Boise ID 83725
vious visits). Individuals from the second group are
then classified as “new” or “recaptured”, depending
on whether they were captured during previous sur-
veys. This visual matching approach works well for
small sets of frogs, but becomes burdensome or even
impossible as the number of frogs captured increases.
The identification problem is to determine from
the photograph if a captured frog is in the existing
database of photographs or is a new frog. Humans
can identify the frogs quite accurately based on the
shape and location of spots or other features on their
skin. For example, in (Lama et al., 2011) the tree frog
Scinax longilineus was successfully identified by re-
searchers simply looking at the collected photographs
and they found that photo-identification was as accu-
rate as tagging the animals. However, as databases of
photographs become large, this visual matching ap-
proach is unrealistic. Instead, researchers are examin-
ing ways to automate this process through computer-
aided pattern recognition.
One of the first steps in pattern recognition is to
identify the area of an animal that will be used for pat-
tern matching. To accomplish this we adopted an ex-
isting tool developed by a research team at Idaho State
University (Vel´asquez, 2006), (Kelly, 2010). An ex-
ample is shown in Figure 1 (Kelly, 2010) which shows
the dorsal (i.e., back) side of the captured frog and in-
dicates the area of its backside which is cutout for use
in the identification. The cutout portion follows natu-
679
Petrovska-Delacrétaz D., Edwards A., Chiasson J., Chollet G. and S. Pilliod D..
Semi-Automated Identification of Leopard Frogs.
DOI: 10.5220/0004828706790686
In Proceedings of the 3rd International Conference on Pattern Recognition Applications and Methods (ICPRAM-2014), pages 679-686
ISBN: 978-989-758-018-5
Copyright
c
 2014 SCITEPRESS (Science and Technology Publications, Lda.)

ral contours of the frog’s backside. This area (referred
to as the region of interest) is then stretched to make a
rectangular array of pixels as shown in Figure 2. The
details of this stretching procedureare givenin (Kelly,
2010).
Ideally, one wants an automatic procedure to iden-
tify the frog, i.e., determine whether or not it is in the
current database. Here we use the terminology semi-
automatic to mean identification of the frog based on
user’s manual selection of the region of interest as in-
dicated in Figure 1. This manual intervention is quite
easy in terms of the user’s effort. In the recognition
approach used in (Kelly, 2010), the cutouts were then
manually segmented to identify the spots and then en-
gineered features were developed for the identifica-
tion procedures. However, the segmentation turned
out to be a rather tedious task to perform on each pho-
tograph.
Figure 1: Cutout along natural contours of the frog (Kelly,
2010).
Figure 2: The cutout is stretched to form a rectangle (Kelly,
2010).
Work similar to that presented here was done by
Gamble et al. (Gamble et al., 2008) who used prin-
cipal component analysis (PCA) on normalized im-
ages of marbled salamanders (Ambystoma opacum).
Specifically, they used a cutout of the back of the
salamander as a vector in R
640×480
and then went
through a series of preprocessing steps to handle nui-
sance variables to obtain M = 625 “new” images for
each original image. Each of these images was then
scaled 8 times (multi-scale in half-octaves from 1 to
8
√
2) using a Gaussian filter and appended to the
original image so that the feature vector was now in
R
9×640×480
. Their database consisted of 366 differ-
ent salamanders and a total of 1008 images. In their
closed-set experiments, they reported a 95% correct
identification that the test image was in the top 10
matches.
In (Azhar et al., 2012) a texture based image fea-
ture descriptor called the Local Binary Patterns (LBP)
was used for the (semi) automatic identification of the
Great Crested Newt salamander (Triturus cristatus).
They tested on a database of 40 newts and 153 images.
Similar to the frog cutout procedure described above,
they used normalized images and manually extracted
a part of the belly images as the source of biometric
information. They considered both open and closed
set test procedures.
The goal of this paper is to provide a simple, fast,
and efficient semi-automated pattern-recognition al-
gorithm for a capture-recapture identification system
for northern leopard frogs (Lithobates pipiens).
Section 2 describes the databases of frog pictures
we used in the experiments. Section 3 discusses how
a PCA algorithm is used to do the animal recogni-
tion (identification), Section 4 describes the evalua-
tion protocols, Section 5 presents the experimental re-
sults and Section 6 gives the conclusions.
2 ANIMAL DATABASE
The database consisted of images of northern leopard
frogs with 209 separate identities. The cutouts of the
frogs described in the introductory section are all rect-
angular arrays of 256×128 pixels (see Figure 2) and
converted to grayscale. This leopard frog database
was provided by the research of Oksana Kelly (Kelly,
2010). Kelly obtained 209 frogs bred in captivity and
photographed them. A photographic light diffusing
dome (Cloud Dome,
www.clouddome.com ) was used
to take an average of 3 to 4 images per frog for all 209
identities, although some frogs had up to 11 images.
The light diffusing dome reduced glare from sunlight,
which helped improveimage quality. Compare Figure
3 with Figure 4. We had 966 images taken with the
dome (hereafter, referred to as Shade Dome images).
There were also 420 additional images taken of frog
identities 109-209 that did not use the shaded dome
(hereafter, referred to as No Dome images). These
images were of significantly lower quality due to glare
(See Figure 4). With the combination of the No Dome
and Shade Dome images, the Captive Leopard frog
database contained 1386 total images.
3 RECOGNITION PROCEDURE
The image capture follows the procedure discussed
in the Introduction. We followed the “finger-
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
680

Figure 3: Photo taken using a shade dome (Kelly, 2010).
Figure 4: Photo taken without using a shade dome (Kelly,
2010).
print” extraction procedure as described in (Kelly,
2010). The open-source program IDENTIFROG
(http://code.google.com/p/identifrog (Pilliod et al., ))
was used to obtain the rectangular cutouts made up
of 256 ×128 pixels as shown in Figure 2. This was
then converted to grayscale for use in the recognition
procedure. We remark that these images of the dor-
sal pattern contained within the fingerprint boundaries
can and do vary in size due to the original image scale
variation, frog positioning, the user’s selection of the
lateral corners of the eyes, and boundary alignment to
the dorsolateral folds.
3.1 Feature Extraction and
Identification
We used Principal Component Analysis (PCA) which
was developed over 100 years ago for statistical anal-
ysis (Pearson, 1901)(Hotelling, 1936). It is also a
well-known method in Machine Learning (Barber,
2012), but for pattern recognition it relies on the im-
ages being normalized. This approach requires deter-
mining a set of images to make up the PCA space (co-
variance matrix), the choice of eigenvectors, and the
choice of an appropriate distance measure. We use
one set of images (development images) to set these
choices and then test on an independent set of images
for the evaluation.
Each frog image x
(k)
is considered to be in R
d
,d ,
256×128 and with N
t
the number of training images,
the covariance of the training set is
C ,
1
N
t
−1
N
t
∑
k=1
(x
(k)
−x
m
)(x
(k)
−x
m
)
T
∈ R
d×d
(1)
x
m
,
1
N
t
N
t
∑
k=1
x
(k)
∈ R
d
. (2)
The rank of C is less than or equal to N
t
−1 and in
our case (typical) N
t
<< d , 256 ×128. As C is a
positive semi-definite symmetric matrix, there is an
orthogonal matrix Q ∈ R
d×d
such that
C = Qdiag(λ
1
,...,λ
N
t
−1
, 0, ...,0
|
{z }
d−(N
t
−1)
)Q
T
. (3)
That is, the i
th
column of Q is the i
th
eigenvector of C
with eigenvalue λ
i
. Further, λ
1
≥ λ
2
≥ ··· ≥ λ
N
t
−1
≥
0. We can then represent any training image x
(k)
∈R
d
as
h
(k)
, Q
T
(x
(k)
−x
m
) ∈ R
d
(4)
since we get the image x back by
x
(k)
= x
m
+ Qh
(k)
∈ R
d
. (5)
However, the point of this approach is to obtain a
compressed representation of the image x by repre-
senting it by its first N eigenvectors where N < N
t
<<
d. That is, the image is coded into R
N
by
h
c
,

h
1
h
2
h
3
··· h
N

T
∈ R
N
(6)
where h
c
is simply the first N components of h ,
Q
T
(x−x
m
) ∈R
d
. With Q
c
∈R
d×N
the first N columns
of Q, the theory of PCA (Barber, 2012) tells us that
reconstruction error is given by
kx−Q
c
h
c
k
2
= λ
2
N+1
+ ···+ λ
2
N
t
−1
. (7)
N is chosen so that λ
2
N+1
+ ···+ λ
2
N
t
−1
is small. Thus,
as far as the Euclidean norm in concerned, the PCA
representation h
c
∈ R
N
is a much lower dimensional
representation of the data than the original data vector
x ∈ R
d
yet provides an accurate reconstruction of the
image.
Let the feature vector be
f =

h
A
··· h
N

T
∈R
N−A+1
(8)
which indicates that we are representing the image by
eigenvectors A through N. Our choice will turn out to
be A = 3,N = 120, that is, we remove the first two
Semi-AutomatedIdentificationofLeopardFrogs
681

components h
1
,h
2
from h
c
to obtain better identifica-
tion accuracy in contrast to reconstruction accuracy.
The basic test is as follows: Let x
(k)
, k = 1, ...,N
t
be the N
t
images in the database and f
(k)
their cor-
responding feature vectors. Let x be any test image
(recapture) with its feature f computed as above. For
k = 1,...,N
t
compute the (cosine of the) minimum an-
gle between the new image and the existing images,
that is, compute
s
(k)
,
f
T
f
(k)
kfk


f
(k)


. (9)
This value s
(k)
is referred to as the score between the
test image and the k
th
image in the database. Let k
∗
be defined by
k
∗
, argmax
k∈{1,...,N
t
}
{s
(k)
} (10)
which we will refer to as the identified image.
In a closed-set protocol the test image of the frog
is assumed to be in the database, such as when a frog
is recaptured during a second sampling event. One
then identifies the test image x as the image x
(k
∗
)
. In
practice one typically finds the (say) 10 images in the
database that score the highest with x and then checks
which of the 10 match the test image.
In an open-set protocol the test image may or
may not correspond to any frog from the reference
database, which is a more realistic test when a frog is
captured and its identity unknown. We again compute
k
∗
as just explained and, with γ some pre-determined
threshold, we check if
s
(k)
≥ γ. (11)
If this is true then x is identified as the image x
(k
∗
)
else we say x is a new identity. Again, in practice,
one typically finds the (say) 10 images in the database
that score closest to the test image and then visually
checks if it matches these already known identities.
Figure 5 shows the mean frog image x
m
and the
first five eigenvectors (eigenfrog images) of the co-
variance matrix constructed using all of the 1386 frog
images (shade dome and no shade dome). Note that
the spots are quite blurred in the first two eigenfrog
images hinting that the first two eigenvectors may not
contribute much to differentiation among frog identi-
ties.
4 EVALUATION PROTOCOLS
We consider both open and closed evaluation proto-
cols. To explain the evaluation protocols, we describe
them for the database consisting of the 209 captive
MeanFrog
50 100 150 200 250
20
40
60
80
100
120
EigF 1
EigF 2
EigF 3
EigF 4
EigF 5
Figure 5: Eigenfrogs.
frog identities with a total of 966 frog images taken
with the shade dome. The number of images for each
frog identity ranged from 2 to 11 with the majority of
the frog identities having 3 to 4 images.
4.1 Closed Set Evaluation Protocol
The closed set protocol assumes that a test frog
is in the database. These frog images are in a
file listed with the number of the frog and the
number of its image. For example,
R 001 01,
R
001 02, R 001 03, R 001 04 are the 4 images
we have of frog 1, R 002 01, R 002 02, R 002 03,
R
002 04, ..., R 002 11 are the 11 images we
have of frog 2, etc. We then distributed the images
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
682

into 5 bins as follows: We put R 001 1 into bin 1,
R 001 2 into bin 2, R 001 3 into bin3, R 001 4 into
bin 4,
R 002 1 into bin 5, R 002 2 into bin 1, R 002 3
into bin 2, R 002 4 , into bin 3, etc. This was done
in order to mix the images of each frog identity well
among the bins. This mixing results in the bin distri-
bution given in Table 1.
Table 1: Bin distribution for the closed set 5-fold protocol.
Bin Number Number of Images
bin 1 194
bin 2 193
bin 3 193
bin 4 193
bin 5 193
After putting all the frog images into the 5 bins
as just described, we used the first four bins to com-
pute the covariance matrix C (PCA subspace). C was
therefore constructed from N
t
= 3×193+ 194 = 773
images. The images in the 5
th
bin were used for test-
ing. We take the identified image to be x
(k
∗
)
where k
∗
is as given in equation (10). The 5
th
bin had 193 im-
ages
3
and three of the images were incorrectly iden-
tified for an accuracy of 190/193 = 98.5%. We then
repeated the procedurefour more times using a differ-
ent bin as the test bin, that is, 5-fold test. The results
are in in Table 2.
Table 2: Closed set 5-fold protocol.
test bin bin 5 bin 4 bin 3
accuracy
190
193
= 98.5%
188
193
= 97.4%
187
193
= 96.8%
test bin bin 2 bin 1
accuracy
187
193
= 96.8%
192
194
= 99%
This shows a total of 22 errors over the fivefolding
(changing the bins five times) tests on the 966 images.
4.2 Open Set Evaluation Protocol
The open set protocol refers to the situation where the
test image (a frog) may or may not be in the database.
In this case we take all 966 of the frog images (209
identities) and separate them into two groups: 804
known frog images with 151 identities and 162 un-
known frog images with 58 identities. The identities
in these two sets are disjoint. As in the closed-set pro-
tocol, the 804 known frog images are distributed into
5 bins of approximately the same size (804/5 giving
161 or 160 images per bin). See Table 3. The covari-
ance matrix (PCA space) is computed using 4 of the
5 bins in the known group. Then the 5
th
bin and the
unknown or 6
th
bin were used to test.
3
966/5 = 193 remainder 1. So 4 of the bins had 193
images and the other bin has 194 images.
Table 3: Bin distribution for the open set protocol.
Bin Number Number of Images
bin 1 161
bin 2 161
bin 3 161
bin 4 161
bin 5 160
bin 6 (unknown frogs) 162
This was repeated a total of five times by per-
muting bins 1 through 5 made up of images from
the known group. The threshold in (11) was chosen
to be γ = 0.5. An error occurs in one of two ways:
(k-known) the test image is in the database, but the
identified image x
(k
∗
)
is not the correct one or (u-
unknown) the test image and its identified image x
(k
∗
)
satisfy the threshold, but the test image in not in the
database. The results are given in Table 4 where u
says the error was made on an unknown frog while k
says that a known frog was misidentified.
Table 4: Open set 5-fold protocol.
test bins bins 5&6 bins 4&6
errors 16 u, 6 k 16 u, 5 k
accuracy
322−22
322
= 0.93%
323−21
323
= 93.5%
test bin bins 3&6 bins 2&6 bins 1&6
errors 11 u, 4 k 14 u, 5 k 15 u, 3 k
accuracy
323−15
323
= 93.4%
323−19
323
= 94.1%
323−18
323
= 94.4%
4.3 Determination of the PCA Space
In using the PCA test we used the feature vector given
in (8) with A = 3,N = 125. That is, we represented
the image as a linear combination of eigenvectors 3
through 125 with the coefficients of this representa-
tion the feature vector. To make this determination we
Figure 6: PCA accuracy as A,N vary.
repeated the closed set evaluation of Subsection 4.1
using eigenvectors A to N where A was varied from 1
to 3 and, for each value of A,N was varied from 50 to
150. The results are shown in the graph of Figure 6.
The graph shows A = 3,N = 125 give good results.
Semi-AutomatedIdentificationofLeopardFrogs
683

5 EXPERIMENTAL RESULTS
Using the open and closed set protocols explained
in the previous section, we performed tests on our
database of images.
5.1 Closed Set of Experiments
5-fold Testing on Shade Dome Images
Consider again the closed set evaluation with the cap-
tive frogs consisting of 209 identities, 966 total im-
ages all taken with the shade dome. We have already
given the results of the first test. However, up to now
we have considered the identified image to be that one
in the database with the highest score [See equations
(9) and (10)]. We now will compare the test image
with the database images with the n highest scores
where we typically take n = 1, 5, or 10. In the present
case we have
Table 5: Shade dome - closed set - 5 fold.
Top n 1 5 10
Avg. errors/fold 4.4 1.6 1.4
Accuracy 97.7% 99.2% 99.3%
For example, in subsection 4.1 we reported the er-
rors for each fold. There were a total of 22 errors in
which the test image did not match the highest score.
This results in average number of errors over the 5
folds given by 22/5 = 4.4. In the case where n = 10,
we see that there were 5×1.4 = 7 times over the five
folds that test image was not one of the top 10 scores.
PCA Constructed from Shade Dome Images
Tested on No Dome Images
In the next experiment we used the 966 shade dome
images to build the PCA space and then tested on the
420 no dome frog images. As previously mentioned,
the no dome images only contained the identities of
frogs 109-209. In Table 6 we report the total number
of misidentification errors.
Table 6: PCA constructed from shade dome images. Tested
on no dome images.
Top n 1 5 10
No. errors 51 31 24
Accuracy 87.9% 92.6% 94.3%
5.2 Open Set Experiments
We next performed an open set experiment using the
966 captive shade dome frogs with 209 identities. As
previously explained in Subsection 4.1, we chose 58
identities with 162 images to be the “unknown” frogs.
The remaining 804 images with 151 identities made
up the “known” frogs. The 804 “known” frogs were
then split into a 5 fold (bins). This was used to make
5 folds (iterations) where for each fold one of the
“known” bins ( ˜161 frogs) along with the 162 “un-
known” frogs were used for testing while the remain-
ing ˜643 frogs were used for development of the PCA
space (covariance matrix). The threshold was set as
γ = 0.5. In Table 7 are the results of only testing the
known frogs and thus is simply the closed set protocol
(γ not used) for the known frogs.
Table 7: Shade dome - open set - 5 fold.
Known frogs - Top n 1 5 10
Avg. errors/fold 2.6 0.4 0.2
Accuracy 98.38% 99.75% 99.88%
However, with the threshold γ = 0.5 it was found
during the 5-fold testing that on average 1.8 of the
known frogs scored below this threshold and thus
would be categorized as a unknown (new) frog.
For the known frogs who met the threshold, Table
8 gives the error results.
Table 8: Known frogs with scores s
(k)
≥ γ = 0.5.
Top n 1 5 10
Avg. Errors/fold 1.4 0.0 0.0
Accuracy 99.13% 100% 100%
With the unknown frogs an error occurs when its
score s
(k)
meets the threshold, i.e., s
(k)
≥ γ = 0.5 be-
cause it is then taken to be in the database when of
course it is not there. Our results are below.
Table 9: Unknown frogs with s
(k)
≥ γ = 0.5.
Top n 1 2 3 4 5
Avg. errors/fold 14.4 5.2 2.4 0.6 0
To explain Table 9, the second column (n = 1)
means that on average there were 14.4 unknown frog
images whose top score with some known frog im-
age was greater than 0.5. The third column (n = 2)
of this table means that on average there were 5.2
unknown frog images whose top 2 scores with some
known frog images were greater than 0.5. Similarly
for the remaining columns. The point here is that if
an unknown frog has a score s
(k)
against the known
database with s
(k)
≥γ = 0.05 then there can be at most
four images in the database that satisfy this threshold
and the biologist need only look at these four images
to visually make the determination that the frog is not
in the database.
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
684

5.3 Threshold
We chose the threshold γ = 0.5. This is based on the
data given in the open set evaluation test discussed
in the previous subsection (Subsection 5.2). Figure
7 shows two probability density functions (pdf). The
pdf in Figure 7 labeled “unknown” frog pdf was ob-
tained by taking the image of each known frog and
computing its score with each unknown frog image.
A histogram of these scores was then normalized to
become the unknown frog pdf.
In contrast, the “known” frog pdf of Figure 7 was
obtained by taking each known frog and computing
its score with itself ( i.e., with all possible images
of its identity) and keeping the highest score. More
precisely, as explained in (the previous) subsection
5.2, the 804 known frogs were put into (essentially) 5
equal sized bins. The PCA space (covariance matrix)
was built from four of the bins. Then each identity in
the remaining (test) bin had its score computed with
images of the same identity in the other 4 bins. The
highest score was kept. This was done five times each
time using a different bin as the test bin. A histogram
made of these scores was normalized to become the
known frog pdf. From the pdfs of Figure 7 the prob-
0 0.2 0.4 0.6 0.8 1
0
2
4
6
8
“unknown” frog pdf
“known” frog pdf
5.0  
5.0
Figure 7: PDF of the scores of known frogs with unknown
frogs and PDF of the scores of known frogs with them-
selves.
ability of detection of a known/unknown frog and the
probability of a false alarm of a known/unknown frog
were computed and are given in Table 10. For exam-
ple, the area under the “unknown” frogs pdf from −∞
to γ = 0.5 is 0.9836 which is the probability that the
score of an unknown frog with a known frog is less
than 0.5. The area under the “known” frogs pdf from
0.5 to ∞ is 0.9838 which is the probability that the
score of a known frog with a known frog is greater
than 0.5.
Of course one should have a separate database for
an evaluation to determine a threshold. However, our
Table 10: Probability of detection.
Test\Truth Unknown Known
Unknown 98.36% 1.62%
Known 1.64% 98.38%
limited amount of images precluded such an opportu-
nity. In practice, the wildlife biologist will not use a
threshold. Instead, the biologist would typically take
a new image and bring up the (say) 10 images in the
database with the top 10 scores. Then a visual check
would be made to determine whether or not the frog
is a new identity. Another way to say this is that the
closed set results matter most to the biologist.
6 CONCLUSIONS
This work was originally motivated by the previ-
ous work of Vel´asquez (Vel´asquez, 2006) and Kelly
(Kelly, 2010). Here we have reported in Table 5 quite
good closed set results which we surmise is due to the
quality of the images. However, table 6 shows that
when “training” (i.e., constructing the PCA space) on
high quality (shade dome) images and then testing on
lower quality (no dome) images the identification ac-
curacy deteriorates. Though the open set results given
in Table 4 are not nearly as good as the closed set re-
sults given in Table 2, Table 9 shows that the biologist
needs to visually check less than five images in the
database to determine if the captured frog is a new
identity or not.
We are in the process of collecting more data to
have enough for both development and evaluation
databases. Another goal is to provide a reference sys-
tem (based on PCA) with a publicly available refer-
ence database, so that the other researchers can com-
pare their results to ours.
ACKNOWLEDGEMENTS
The authors gratefully acknowledge the Department
of Electrical and Computer Engineering at Boise State
University for providing funding for this project.
REFERENCES
Azhar, M., Hoque, S., and Deravi, F. (2012). Automatic
identification of wildlife using local binary patterns.
In IET Conference on Image Processing (IPR 2012).
London UK.
Barber, D. (2012). Bayesian Reasoning and Machine
Learning. Cambridge.
Semi-AutomatedIdentificationofLeopardFrogs
685

Gamble, L., Ravela, S., and McGarigal, K. (2008). Multi-
scale features for identifying individuals in large bi-
ological databases: An application of pattern recog-
nition technology to the marbled salamander Am-
bystoma Opacum. Journal of Applied Ecology,
45:170–1180.
Hotelling, H. (1936). Relations between two sets of vari-
ates. Biometrika, pages 321–377.
Kelly, O. V. (2010). Automated Digital Individual Identi-
fication System with an Application to the Northern
Leopard Frog Lithobates pipiens. PhD thesis, Idaho
State University.
Lama, F. D., Roca, M. D., Andrade, M. A., and Nascimento,
L. B. (2011). The use of photography to identify indi-
vidual tree frogs by their natural marks. South Ameri-
can Journal of Herpetology, 6(3):198–204.
Pearson, K. (1901). On lines and planes of closest fit to
systems of points in space. Philosophical Magazine,
2(11):559–572.
Pilliod, D., Velasquez, E., Bosworth, K., Ahsan,
H., and Kelly, O. Identifrog: An automated
pattern recognition program for leopard frogs.
http://code.google.com/p/identifrog/.
Vel´asquez, M. E. (2006). Wavelets: Theory and Applica-
tions. PhD thesis, Idaho State University, Pocatello,
Idaho, USA.
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
686