Journal of Mammalogy, 91(6):1350–1359, 2010

Computer-aided photo-identification system with an application to

polar bears based on whisker spot patterns
CARLOS J. R. ANDERSON,* NIELS DA VITORIA LOBO, JAMES D. ROTH, AND JANE M. WATERMAN
Department of Biology, University of Central Florida, Orlando, FL 32816, USA (CJRA, JDR, JMW)
Department of Computer Science, University of Central Florida, Orlando, FL 32816, USA (NDVL)
* Correspondent: carlosja@msu.edu

Downloaded from https://academic.oup.com/jmammal/article/91/6/1350/888329 by guest on 29 February 2024

Ecologists often rely on unique natural markings to identify individual free-ranging animals without disturbing
them. We developed a computer-aided photo-identification system for identifying polar bears (Ursus
maritimus) based on whisker spot pattern recognition. We automated our system so that the selection of 3
reference points on the input image is the only manual step required during image preprocessing. Our pattern-
matching algorithm is unique in that the variability within spot patterns is considered fully rather than
representing them as points and applying a point-pattern matching algorithm. We also measured the reliability
of our method as probabilities of true positives and false positives using photographs of various qualities taken
at different angles. When we excluded photographs of poor quality and angle the probability of true positives
was .80% at a false positive probability of 10%. A new photograph could be preprocessed in ,1 min and
tested against a reference library of 100 individuals in ,10 min. Our computer-aided identification system
could be extended for use in other species with variable spot patterns, which could be useful in efforts to
estimate various population dynamics parameters essential for the study and conservation of wildlife,
particularly threatened and endangered species. DOI: 10.1644/09-MAMM-A-425.1.

Key words: capture–recapture, Chamfer, computer vision, image pattern matching, noninvasive identification, reliability,
Ursus maritimus

E 2010 American Society of Mammalogists

In studies of population dynamics using capture–recapture
models, estimates of population size and rates of survival,
reproduction, and migration can be obtained only if animals
can be identified uniquely (Nichols 1992). Because animals
often differ in their individual behavior (Fagen and Fagen
1996; Martin and Kraemer 1987), individual recognition of
animals is also essential in studies of behavioral ecology. In
analyses of animal movement patterns individual animals are
frequently tracked through space and time (Parks et al. 2006).
Thus, many kinds of ecological studies involve identifying
individual animals in the field (Nietfeld et al. 1994).
One method of identifying animals individually is by
applying artificial markings, such as tags, tattoos, or tissue
removal (Nietfeld et al. 1994). Artificial markings are unique
for each animal, making this method very reliable, unless the
animal loses its markings (Diefenbach and Alt 1998).
Applying artificial markings, however, is invasive because
the animal typically must be captured and handled to be
marked and often recaptured to be identified (Pennycuick and
Rudnai 1970). Consequently, this method could be difficult
and expensive to use and possibly detrimental to the health or
behavior of the animal (Kelly 2001; Pennycuick 1978). For
1350
threatened or endangered species this method can be
especially undesirable or even prohibited (Kelly 2001;
Pennycuick and Rudnai 1970). Additionally, visible tags or
tattoos can interfere with recreational or professional wildlife
photography. Marking animals can be convenient, however,
when handling is necessary for measuring physiological traits,
sampling blood or tissue, or attaching radiocollars for tracking
(Ramsay and Stirling 1986).
An alternative method of identifying animals individually is
through natural markings, such as pelage spot patterns on
cheetahs (Acinonyx jubatus—Kelly 2001), fin marks on
cetaceans (Hammond et al. 1990), and whisker spot patterns
on lions (Panthera leo—Pennycuick and Rudnai 1970) or
polar bears (Ursus maritimus—Anderson et al. 2007). The
main advantage of using natural markings is that it is generally
noninvasive because the animal can be identified from a
distance without disturbing it. In addition, using natural
markings is practical and affordable because it involves
operating accessible equipment (e.g., a photographic camera).
www.mammalogy.org
December 2010 ANDERSON ET AL.—AUTOMATED IDENTIFICATION OF POLAR BEARS
The main disadvantage, however, is that unless the entire
study population is sampled, it is not possible to know for
certain whether natural markings of every individual are
unique (Pennycuick 1978). In addition, natural markings can
be difficult to distinguish from a distance, or might be lacking
altogether, which can lead to incorrect identification (Oliveira-
Santos et al. 2010). Furthermore, searching for an individual’s
identity by visually comparing hundreds or thousands of
natural patterns can be tedious, error-prone, and time-
consuming (Hillman et al. 2003; Kelly 2001).
To overcome these disadvantages various computer-aided
identification systems have been developed to recognize and
match natural markings. Recognition of pelage or skin spot
patterns, for example, has been used to identify individual
seals (Hastings et al. 2008; Hiby and Lovell 1990), cheetahs
(Kelly 2001), whale sharks (Rhincodon typus—Arzoumanian
et al. 2005), and spotted raggedtooth sharks (Carcharias
taurus—van Tienhoven et al. 2007). Sperm whales (Physeter
catodon—Huele and de Haes 1998; Whitehead 1990) and
many other marine animals have been identified uniquely by
computerized recognition of edge patterns on their fins or
flukes (Hillman et al. 2003). Stripes or bands have been used
by computer-aided identification systems, most recently in
zebras (Equus burchellii—Foster et al. 2007) and marbled
salamanders (Ambystoma opacum—Gamble et al. 2008).
In many computer-aided identification systems each new
image required substantial preprocessing (e.g., cropping or
enhancing brightness and contrast), which can be time-
consuming, require training, and introduce subjectivity and
error (Kelly 2001). Furthermore, individual features (e.g.,
spots and stripes) of natural patterns often vary in shape and
size, yet many computer-aided identification systems ignore
this variability and focus on the arrangement of these features
instead. Some identification systems use ‘‘blob extraction’’ to
translate a pattern into a set of (x, y) points, to then apply a
point pattern matching algorithm to compare a pair of patterns.
This method works well for features with little variation in
shape or size (Arzoumanian et al. 2005; van Tienhoven et al.
2007) or when the arrangement of the features composing the
pattern is more important than the shape or size of the features
themselves. When the individual features of a pattern are
important, however, translating these features to simpler
representations can reduce the reliably of identification.
Moreover, irregular or complex natural patterns cannot be
translated easily to a simpler representation.
Computer-aided identification systems often work by
comparing the photograph of the animal in question with the
photographs of known individuals (i.e., a reference library).
For each comparison of the unknown photograph with a
reference photograph, the system produces a numerical
similarity score. If the similarity score is below—or above,
depending on the system—a specified similarity threshold, the
photographs are considered a potential match (i.e., they likely
belong to the same individual). Similarity scores often are
ranked to place the best scores 1st, making it easier for the
user to evaluate potential matches.
1351
A reliable identification system should produce similarity
scores for photographs of the same individual that are well
below the similarity threshold (in this study, we assume that
the lower the score, the better the match, but this criterion
depends on the specific identification system). Two main
types of error, however, degrade the reliability of an
identification system: photographs of different individuals
are incorrectly recognized as a match (i.e., a false positive);
and photographs of the same individual are not recognized as a
match (i.e., a false negative—Hastings et al. 2008; Kelly
2001). The 1st type of error, or probability of false positives,
Downloaded from https://academic.oup.com/jmammal/article/91/6/1350/888329 by guest on 29 February 2024
reflects the proportion of individuals in the reference library
that must be evaluated by the user for every identification
attempt. For example, if a reference library contains 100
individuals and the probability of false positives is 10%, every
attempt would list (on average) 10 incorrect individuals below
the similarity threshold. The 2nd type of error commonly is
assessed not by the probability of false negatives but by the
probability of true positives (12probability of false nega-
tives), or simply the probability of obtaining a correct match
between the same individuals.
Both types of error are controlled by the similarity threshold
simultaneously. For example, the probability of true positives
can be made as high as desired by modifying the similarity
threshold. The trade-off, however, is that the higher the
probability of true positives, the higher the probability of false
positives. To evaluate this trade-off the assessment of an
identification system should include both its probability of
true positives and probability of false positives. Often,
however, only the probability of true positives is reported,
which not only makes it difficult to compare performance
among various systems but also obscures the amount of
additional user effort required in evaluating potential matches.
We developed and evaluated a novel computer-aided photo-
identification system for animals with distinct natural patterns
and applied it to polar bears using their whisker spot patterns.
Polar bear whisker spots are found at the base of the whisker
follicles, arranged in a distinct pattern on each side of the
anterior end of the muzzle, and can be used to identify
individual bears reliably (Anderson et al. 2007). Because the
size and shape of these spots can vary, we developed a unique
matching algorithm that operates on the image itself without
translating the pattern into an alternative representation. We
determined the trade-off between the probability of true
positives and the probability of false positives in our system
and evaluated its performance against photographs of different
quality taken at various angles. To automate the matching
process as much as possible, minimizing user involvement, the
only steps in our system involving the user are choosing 3
reference locations on an input image and verifying that
putative matches are correct.
MATERIALS AND METHODS
Study site and collection of polar bear photographs.—We
photographed polar bears approximately 30 km east of the
1352 JOURNAL OF MAMMALOGY Vol. 91, No. 6

town of Churchill, Manitoba, Canada (58u459N, 93u459W).

Polar bears congregate along the western shore of Hudson Bay
as the sea ice melts in August and remain on the coast until
freeze-up in mid-November (Latour 1981). We accessed our
study site via a tundra vehicle, a large bus adapted for
travelling on tundra and normally used for ecotourism (Dyck
and Baydack 2004). No more than 18 tundra vehicles were
permitted within our site. Polar bears rarely responded to the
approach of a vehicle (Eckhardt 2005) and were free to leave
the site at any time. All procedures were in accordance the
guidelines of the American Society of Mammalogists (Gannon

Downloaded from https://academic.oup.com/jmammal/article/91/6/1350/888329 by guest on 29 February 2024

et al. 2007) and were approved by the University of Central
Florida Animal Care Committee (04-34W).
Photographs of polar bears were taken daily (0900–1500 h)
by trained volunteers and ourselves during 18 October–11
November 2003, 18 October–10 November 2004, and 18
October–10 November 2005, using Nikon D100 6.0-mega-
pixel digital cameras and 70–300 and 80–400 mm lenses
(Nikon, Melville, New York). When possible, polar bears
were identified individually by facial scars, sex, and body
shape and size (Eckhardt 2005; Eckhardt et al. 2002).
Photographs of the same polar bear were taken multiple times
throughout each field season. We used photographs of the
same polar bear taken 1 day apart (in some cases, 1 year
apart) to evaluate our system.
Computer-aided identification system.—Our identification
system consists of 3 main components: the reference library,
the image preprocessing method, and the matching algorithm.
The reference library stores images of known individuals and
is used as the source of images to match new animals. The
image preprocessing method automatically extracts the natural
pattern of interest from an image by standardizing and
enhancing the image. Finally, the matching algorithm
computes the similarity score between 2 images. The user
must select 3 locations, or reference points, on an input image
for the system to orient the image automatically and find the
whisker spot pattern region. For polar bears these points are FIG. 1.—a) Polar bear photograph marked by the user with 3
the front corner of the eye, the notch of the nose, and the reference points required by the identification system: 1) front corner
trailing edge of the mouth (Fig. 1a). Our identification system of the eye, 2) notch of the nose, and 3) trailing edge of the mouth. b)
Standardization of the photograph by an affine geometric transform.
was written in Microsoft Visual C# 2005 Express Edition
(.NET Framework 2.0; Microsoft Corp., Redmond, Washing-
ton). All computer development and analyses were conducted on pixels to the right of the corner of the eye and the edge of the
a Latitude D620 (Dell, Round Rock, Texas) with dual 2.0 GHz mouth lay 128 pixels to the right and 128 pixels below the eye
processors and 2.0 GB of memory. A summary of the steps taken (Fig. 1b). The resulting image was small enough to keep the
by our polar bear identification system is found in Appendix I. computational time reasonable while allowing small whisker
Image preprocessing.—The input image was 1st converted spots to remain visible. Standardizing the input image to a
to grayscale using the luminance formula 0.30 3 R + 0.59 3 G predefined orientation and size facilitated comparing it with
+ 0.11 3 B (Bunks 2000), where R, G, and B are the red, green, the reference library images, which also were standardized.
and blue values of a pixel. Grayscale conversion simplified Next, the image was cropped around the region where
the implementation of the methods used below without whisker spots typically are found (Fig. 2a). The cropped
sacrificing information loss, given that whisker spot regions image then was enhanced 1st by histogram matching
of polar bears are not colorful. With use of the 3 reference (Gonzalez and Woods 2002), which adjusts the histogram of
points selected by the user, the image was geometrically pixel values of the image to match the histogram of a good-
(affine-) transformed (Foley et al. 1995) to a predefined quality image of polar bear whisker spots; and 2nd, by
orientation and size using the Graphics.Transform property in logarithmic transformation (Fisher et al. 2004; Gonzalez and
Visual C# (Petzold 2002) so that the notch of the nose lay 240 Woods 2002), which applies a logarithmic function to every
December 2010 ANDERSON ET AL.—AUTOMATED IDENTIFICATION OF POLAR BEARS
FIG. 2.—Preprocessing of the input photograph of a polar bear. a)
The image was cropped automatically around a fixed area where
whisker spots are typically found, b) enhanced by histogram
matching and logarithmic transformation, c) smoothed by neighbor-
hood averaging, and d) turned to black and white by applying
adaptive thresholding repeatedly.
pixel value of the image to improve its contrast (Fig. 2b). The
enhanced image was smoothed by neighborhood averaging
(Gonzalez and Woods 2002) with a neighborhood radius of 2
to remove noise (Fig. 2c).
The smoothed image was converted to black and white (not
grayscale) using adaptive thresholding (Davies 2004; Foster
et al. 2007), which changes to black any pixel whose value is
lower than a threshold relative to its neighborhood—
otherwise, the pixel is changed to white. We found, however,
that using a predefined threshold value for adaptive thresh-
olding did not preserve the original size of whisker spots on
low-quality images. Thus, the threshold value for a particular
image was determined by testing adaptive thresholding with
multiple, increasing threshold values (starting with a value of
2) until the total number of black pixels on the entire image
was 300. Because the total number of black pixels on an
image decreases each time the threshold value is increased,
our condition of the upper limit on the number of black pixels
always will be met. This limit value on the number of black
pixels and the size of the neighborhood (radius of 4 pixels)
were determined empirically by testing this algorithm with
several preprocessed high-quality images. Finally, the result-
ing image was cropped again to eliminate extraneous artifacts
around the edges left by adaptive thresholding (Fig. 2d).
Computation of the similarity score.—We used the Chamfer
distance algorithm (Borgefors 1986) to compute the similarity
score between 2 preprocessed images. The Chamfer distance
algorithm looks at each black pixel of the 1st image and
determines its Euclidean distance to the nearest black pixel of
1353
the 2nd image, and uses the mean of these distances as the
final distance between the images. Note that ‘‘nearest’’ does
not imply nearby because the nearest black pixel of the 2nd
image could be relatively far from the 1st and therefore will
have a large Euclidean distance that will increase the overall
Chamfer distance between the images. Also, note that if 2
black pixels are closely aligned, the distance used in the
Chamfer calculation is very small (0 for a perfect alignment),
regardless of the presence of other black pixels. However,
when 2 images come from the same individual but 1 of the
images contains spurious pixels (i.e., black pixels that are not
Downloaded from https://academic.oup.com/jmammal/article/91/6/1350/888329 by guest on 29 February 2024
true identifying spots), the distance between each spurious
pixel and the nearest black pixel of the other image will be
large. These large distances will cause the overall Chamfer
distance to be large even though the images come from the
same individual.
To reduce the influence that spurious pixels had on the
resulting similarity score, our implementation used the median
of distances (rather than the mean) to calculate the Chamfer
distance because the median of a set of numbers is less
affected by outliers than the mean. Because the Chamfer
distance differs depending upon which image is 1st or 2nd, our
algorithm also calculated the Chamfer distance of the images
in reverse order and averaged these 2 distances (Fig. 3).
Furthermore, to account for image misalignments caused by
foreshortening (i.e., rotations toward or away from the
camera), our algorithm calculated the Chamfer distance
repeatedly while moving one image about the other up to 16
pixels in each direction (in 2-pixel increments for faster
computation, but this increment value can be changed by the
user), and used the minimum of these distances as the final
similarity score (Fig. 4). The affine transformation applied
during image preprocessing best handled clockwise or
counterclockwise rotations, but the shifting procedure above
handled rotations due to foreshortening.
Reliability of the identification system.—Photographs of the
same polar bear were chosen so that they were taken 1 day
apart (in some cases, 1 year apart). Photographs were
categorized according to angle (excellent, moderate, and poor)
and quality (excellent, moderate, and poor). Angle was based on
how perpendicular the facial profile of a polar bear was to the
camera axis: excellent deviated ,15u from the camera axis,
moderate deviated 15–30u, and poor deviated 30–45u. Photo-
graphs with angle .45u were not used in our study because
whisker spots were imperceptible to us. We estimated these
angle deviations using a miniature polar bear model rotated to
match the perspective of the polar bear in the photograph.
Quality was estimated based on image sharpness (Fig. 5).
We used our identification system to compute the similarity
scores for pairs of photographs of the same polar bear with
excellent angle and quality. Then, to examine the effect of
photographic angle and quality we computed the scores for
pairs of photographs of the same bear, with one photograph
having excellent angle and quality and the other having either
medium angle and excellent quality or excellent quality and
medium angle. Also, we computed the scores for photographs
1354 JOURNAL OF MAMMALOGY
FIG. 3.—Sample images A and B overlapped to illustrate how the
similarity score between 2 images was calculated. Each square
corresponds to a pixel of either image. The number on each black
pixel is the Euclidean distance, rounded to the nearest 0.1, between
the center of that pixel and the center of the nearest black pixel on the
other image. The distance of image A to B is 1.0 (median of numbers
on image A). The distance of image B to A is 1.2 (median of numbers
on image B). Thus, the similarity score between images A and B is 1.1
(average of above distances).
of different polar bears with excellent angle and quality to
examine the relationship between the probability of true
positives and the probability of false positives. As an
additional measure of reliability we sorted the similarity
scores in ascending order (i.e., best to worst) of 37 matching
attempts using photographs of excellent and moderate quality
and angle. We recorded the position in the sorted list of the
correct match, which would indicate the number of matches in
our identification system that would have to be verified before
encountering the correct match.
RESULTS
From .200 individuals photographed we found 57
individuals with photographs of excellent angle and quality,
resulting in 796 possible comparisons of the same side of the
face of different polar bears (25 photographs of the left side
and 32 of the right side). Of these 57 individuals, 8 had a 2nd
photograph of the same side of the face of excellent angle and
Vol. 91, No. 6
quality, 19 had another photograph of excellent angle and
moderate quality, 14 had another photograph of excellent
angle and poor quality, 18 had another photograph of
moderate angle and excellent quality, and 18 had another
photograph of poor angle and excellent quality.
The mean similarity score for comparisons of different
photographs of the same polar bear increased (i.e., photo-
graphs were less alike) as the level of quality or angle of one
of the photographs decreased from excellent to moderate to
poor (Fig. 6). The mean score was highest, however, for
comparisons of photographs of different polar bears (Fig. 6).
Downloaded from https://academic.oup.com/jmammal/article/91/6/1350/888329 by guest on 29 February 2024
As we increased the similarity threshold the proportion of
comparisons between different polar bears with scores below
that threshold (i.e., the probability of false positives) also
increased (Fig. 7).
The probability of true positives increased as the probability
of false positives increased (Fig. 8). The probability of true
positives decreased, however, when we used photographs of
moderate angle or quality and decreased further with
photographs of poor angle or quality (Fig. 8). For example,
the probability of true positives using photographs of moderate
angle was only 80% when the probability of false positives
was 5%, but the probability of true positives increased to 90%
when the probability of false positives was 10%. When we
used photographs of excellent quality and angle only (albeit a
small sample), the probability of true positives was 100%
when the probability of false positives was only 2%.
When the similarity scores of 37 matching attempts using
photographs of excellent and moderate angle and quality were
sorted from lowest to highest (i.e., best to worst), the correct
match was listed first 27 (73%) times (Table 1). This sorted
list of scores is how our identification system reports all
matching results, such that the user is always able to verify
matches (or mismatches) by eye. The overall mean position of
the correct match in the sorted list was 2.0 6 0.4 SE (n 5 37)
in a reference library of either 25 (left-sided photographs) or
32 (right-sided photographs).
DISCUSSION
Our computer-aided identification system, which includes
information on the size and shape of whisker spots and their
location, proved reliable when applied to polar bear whisker
spots. Our system had a reasonable probability of true
positives (80%) with a low probability of false positives
(10%) when we excluded photographs of poor angle and
quality. At this false positive probability about 3 potential
matches using a reference library of 30 individuals would have
to be verified by the user to obtain an expected probability of
true positives of 80%. We found the correct match 86.5% of
the time when we verified the top 3 potential matches when
similarity scores were ranked (Table 1). The probability of a
true positive improved when similarity scores were sorted
because correct matches often had lower scores than the other
potential matches, even if the scores of correct matches were
above the similarity threshold.
December 2010 ANDERSON ET AL.—AUTOMATED IDENTIFICATION OF POLAR BEARS 1355

Downloaded from https://academic.oup.com/jmammal/article/91/6/1350/888329 by guest on 29 February 2024

FIG. 4.—Surface plots of similarity scores (0 to 8) determined from 2 images, shifted up to 16 pixels in each direction (horizontal and
vertical). Each column of plots corresponds to comparisons of images with various photographic angles: 1) excellent angle (EA) for both images
of the same polar bear, 2) 1 excellent and 1 moderate angle (MA) image, 3) 1 excellent and 1 poor angle (PA) image, and 4) excellent angle for
both images of different polar bears. No polar bear was used more than once for any comparison in this figure.

Poor-angle photographs caused our system to misalign the
images while computing the similarity score because polar bear
whisker spot patterns were foreshortened. In a computer-aided
identification system for whale sharks (R. typus) based on skin
spot patterns, foreshortening also produced mismatches when
using photographs with angle .30u (Arzoumanian et al. 2005).
Nevertheless, using poor-angle photographs is useful because
having a record of an individual at various angles improves its
chance of being recognized in the future (Arzoumanian et al.
2005; Hillman et al. 2003; Kelly 2001).
Polar bear photographs with poor quality caused smaller
whisker spots to disappear during preprocessing, effectively
changing the extracted pattern. Our method of repeatedly
applying adaptive thresholding to improve robustness was
most effective with moderate-quality photographs. Objective
definitions of photographic quality are difficult to compare
across studies, but others also have found that using poor-
quality photographs reduces the probability of true positives
(Kelly 2001; Whitehead 1990).
The appropriate similarity threshold for our system can be
chosen based on the trade-off between the probability of true
positives and probability of false positives. For example, we
might tolerate a 20% probability of false positives if our
reference library was small—for example, 100 individuals—
because the user would have to verify only about 20 individuals
on average before finding the correct one. In contrast, if our
database was large—for example, 1,000 individuals—we might
tolerate a 5% probability of false positives at the cost of a lower
probability of true positives. We then can use the relationship
between the similarity threshold and the probability of false
positives to determine the similarity threshold we need to use.
We recommend that future evaluations of identification systems
show the relationship between the probability of true positives
and probability of false positives, or at least report the
probability of false positives alongside the reported probability
of true positives at the chosen similarity threshold.
We found that most excellent- and moderate-quality
photographs were taken at a distance ,50 m from the polar
bears. Anderson et al. (2007) found that whisker spots of polar
bears were most distinguishable in photographs taken ,50 m
away from the bears. Future improvements in digital
photography and more advanced optical equipment should
allow more-distant animals to be recognized. The current
distance limitation might preclude the use of our system in
remote locations where achieving proximity to polar bears is
difficult (Anderson et al. 2007).
In species lacking stripes or spots authors often argue that
they can use subtle marks, scars, and other attributes to
1356 JOURNAL OF MAMMALOGY
FIG. 5.—Photographs of different polar bears were manually
categorized as a) excellent quality, b) and c) moderate quality, and d)
poor quality. Excellent-quality photographs were well focused and
crisp, clearly showing individual whisker spots. Moderate-quality
photographs lacked sharpness, but individual whisker spots could still
be discerned. Poor-quality photographs were unfocused and blurry,
making it difficult to distinguish individual whisker spots.
identify individuals reliably (Oliveira-Santos et al. 2010).
Oliveira-Santos et al. (2010) tested the ability of researchers to
discriminate among individuals in a species without clear
identifying marks. They sent 55 photographs of 8 individual
FIG. 6.—Mean similarity score for comparisons of different
photographs of the same polar bear (first 2 sets of bars) and
photographs of different bears. The score for comparisons of the same
bears increased (i.e., decreased in similarity) as the quality or angle of
one of the photographs decreased from excellent to moderate to poor
(the other photograph in each comparison had excellent quality and
angle). The mean score was highest when we compared photographs
of different bears (both photographs in each comparison had excellent
quality and angle). Error bars represent 1 SE. Numbers above bars
represent the number of comparisons in each category.
Vol. 91, No. 6
Downloaded from https://academic.oup.com/jmammal/article/91/6/1350/888329 by guest on 29 February 2024
FIG. 7.—Relationship between the similarity threshold and the
probability of a false positive for comparisons of polar bear
photographs. The higher the threshold is set, the more comparisons
with a similarity score below the threshold (i.e., incorrect matches).
lowland tapirs (Tapirus terrestris) to 14 researchers for
identification and found that errors in identification ranged
up to 75% even though they used a small sample size of
individuals and photographs. They concluded that haphazard
identification of individuals from photographs can lead to
numerous misidentifications. Like tapirs, polar bears have
little in the way of external identifying marks to use to
discriminate among individuals. In our research we are
investigating the behavior of polar bears in a population of
approximately 1,000 animals (Regehr et al. 2007), more
individuals than we could possibly match by eye based on
natural marks. The use of whisker spots is a reliable
discriminator among individuals for this species (Anderson
et al. 2007), but whisker spots vary in position, number, and
FIG. 8.—Relationship between the probability of a false positive
and the system accuracy for photographs of excellent, moderate, and
poor quality and angle. In general, as the probability of false positives
increased the system accuracy increased. For any specific probability
of false positives the system accuracy decreased with decreasing
quality or angle.
December 2010 ANDERSON ET AL.—AUTOMATED IDENTIFICATION OF POLAR BEARS
TABLE 1.—Number of matching attempts (from a total of 37) that
placed the correct match between 2 polar bear photographs (of
excellent and moderate angle and quality) at the specified position in
the sorted list of similarity scores.
Position of correct match No. matching attempts
1 27
2 4
3 1
5 1
6 2
7 1
13 1
size. The combination of patterns to discern individuals is too
complicated for us to determine, 1st, individual identity by eye
in a sample of 37, or even as few as 20, bears (it would take
more than a day to do 100 individuals), and 2nd, if an
individual is new and not one from our known catalog of
bears. Our computer program, on a laptop, would allow us to
narrow down possible matches to 1 or 2 individuals and thus
avoid the problems found by Oliveira-Santos et al. (2010).
Our identification system took up to 5 s to compute the
similarity score for a single image comparison using a Latitude
D620 laptop with dual 2.0 GHz processors and 2.0 GB of
memory. Thus, our system would take ,10 min to match 1
photograph against a reference library of 100 individuals (i.e.,
length of time grows linearly with the size of the reference
library). Because we often observed no more than 10 new polar
bears in the field per day, our system is fast enough to be used in
the field (i.e., on a laptop). Inputting a photograph into the
system took ,1 min, including selecting the 3 reference points
and the time the system took to enhance, or preprocess, the
image automatically. In contrast, some identification systems
require the user to preprocess images or perform additional
steps manually, which can take several more minutes.
The matching algorithm we used to compute the similarity
score of a pair of images was based on the Chamfer distance
algorithm, a computer vision technique that works on images
directly. Other systems use ‘‘blob extraction’’ to transform a
pattern of spots into a set of coordinates and then apply a point
pattern matching algorithm to compute the similarity score,
which works well for spots with regular shapes (Arzoumanian
et al. 2005). Because polar bear whisker spots can vary in
shape and size, however, we applied an algorithm that
captured this variability using image pattern matching rather
than simplifying the pattern into dimensionless points. Our
unique method of matching is especially useful in recognizing
species that have natural patterns other than spots (e.g., stripes,
rings, and irregular blotches). Gamble et al. (2008) also
considered the shape and size of features through direct
comparison of images and obtained excellent results (95%
true positives with 1% false positives), although their system
took 4–5 min per image comparison whereas our system took
only about 5 s per comparison.
The success of our identification system depended on the
reliability of whisker spot patterns, which is high in polar bears
1357
(Anderson et al. 2007). Adult polar bear whisker spot patterns
change little from year to year, but it is unknown whether
whisker spot patterns change with the bear’s maturation or
whether they are similar among related polar bears (Anderson et
al. 2007). In African penguins (Spheniscus demersus), chest
spots on juveniles are not reliable (T. Burghardt, University of
Bristol, pers. comm.), and thus only adults could be identified
using an identification system. Kelly (2001) found no difference
between similarity scores of related cheetahs (A. jubatus) and
scores of unrelated cheetahs, implying that cheetah spot patterns
were not similar among related individuals.
Downloaded from https://academic.oup.com/jmammal/article/91/6/1350/888329 by guest on 29 February 2024
We already have used our identification system in studies of
polar bear behavior in western Hudson Bay. With our system
we have identified several polar bears that were previously
unknown due of lack of distinct markings (other than whisker
spots). Our identification system can be modified for use with
other species where invasive identification can be difficult or
undesirable. For example, we recently modified our system to
recognize individual boreal toads (Bufo boreas boreas) from
their irregular belly markings (in collaboration with K.
Thompson, Colorado Division of Wildlife, Montrose, Colora-
do), where we have had good preliminary results. Our system
is especially ideal for identifying animals with irregular
markings because our algorithm considers the shape and size
of individual spots or other features. Finally, our system could
be used in conjunction with capture–recapture methods to
estimate ecological parameters essential for the study,
management, and conservation of animals.
ACKNOWLEDGMENTS
Funding was provided by The Earthwatch Institute, Polar Bears
International, and the Cotswold Foundation. We are grateful to the
many volunteers who helped us in the field (especially M. Miller, B.
Pettitt, and A. Wubbels), our tundra buggy driver B. Debits, and the
staff of the Churchill Northern Studies Centre. We thank R.
Buchanan and C. Buchanan for their generous support. We are
grateful to C. Schwarz for his programming assistance and K. Osborn
for providing additional literature. We thank K. Thompson for his
interest and involvement in applying our system to identifying boreal
toads. Finally, we thank J. Holmberg for providing valuable
comments that helped improve the quality of this paper.
LITERATURE CITED
ANDERSON, C. J. R., J. D. ROTH, AND J. M. WATERMAN. 2007. Can
whisker spot patterns be used to identify individual polar bears?
Journal of Zoology (London) 273:333–339.
ARZOUMANIAN, Z., J. HOLMBERG, AND B. NORMAN. 2005. An
astronomical pattern-matching algorithm for computer-aided
identification of whale sharks Rhincodon typus. Journal of Applied
Ecology 42:999–1011.
BORGEFORS, G. 1986. Distance transformations in digital images.
Computer Vision, Graphics, and Image Processing 34:344–371.
BUNKS, C. 2000. Grokking the GIMP. New Riders Publishing,
Thousand Oaks, California.
DAVIES, E. R. 2004. Machine vision: theory, algorithms, practicalities.
3rd ed. Morgan Kaufmann Publishers Inc., San Francisco,
California.
1358 JOURNAL OF MAMMALOGY
DIEFENBACH, D. R., AND G. L. ALT. 1998. Modeling and evaluation of
ear tag loss in black bears. Journal of Wildlife Management
62:1292–1300.
DYCK, M. G., AND R. K. BAYDACK. 2004. Vigilance behaviour of polar
bears (Ursus maritimus) in the context of wildlife-viewing
activities at Churchill, Manitoba, Canada. Biological Conservation
116:343–350.
ECKHARDT, G. H. 2005. The effects of ecotourism on polar bear
behavior. M.S. thesis, University of Central Florida, Orlando.
ECKHARDT, G. H., J. M. WATERMAN, AND J. D. ROTH. 2002. The
functional significance of play fighting in polar bears: are they
asocial? Integrative and Comparative Biology 42:1224.
FAGEN, R., AND J. M. FAGEN. 1996. Individual distinctiveness in brown
bears, Ursus arctos L. Ethology 102:212–226.
FISHER, R., S. PERKINS, A. WALKER, AND E. WOLFART. 2004.
Hypermedia image processing reference. http://homepages.inf.ed.
ac.uk/rbf/HIPR2. Accessed 15 May 2010.
FOLEY, J. D., A. VAN DAM, S. K. FEINER, AND J. F. HUGHES. 1995.
Computer graphics: principles and practice. 2nd ed. Addison-
Wesley, Reading, Massachusetts.
FOSTER, G., H. KRIJGER, AND S. BANGAY. 2007. Zebra fingerprints:
towards a computer-aided identification system for individual
zebra. African Journal of Ecology 45:225–227.
GAMBLE, L., S. RAVELA, AND K. MCGARIGAL. 2008. Multi-scale
features for identifying individuals in large biological databases: an
application of pattern recognition technology to the marbled
salamander Ambystoma opacum. Journal of Applied Ecology
45:170–180.
GANNON, W. L., R. S. SIKES, AND THE ANIMAL CARE AND USE COMMITTEE
OF THE AMERICAN SOCIETY OF MAMMALOGISTS. 2007. Guidelines of
the American Society of Mammalogists for the use of wild
mammals in research. Journal of Mammalogy 88:809–823.
GONZALEZ, R. C., AND R. E. WOODS. 2002. Digital image processing.
2nd ed. Prentice Hall, Upper Saddle River, New Jersey.
HAMMOND, P. S., S. A. MIZROCH, AND G. P. DONOVAN. 1990. Individual
recognition of cetaceans: use of photo-identification and other
techniques to estimate population parameters. International
Whaling Commission, Cambridge, Massachusetts.
HASTINGS, K. K., L. A. HIBY, AND R. J. SMALL. 2008. Evaluation of a
computer-assisted photograph-matching system to monitor natu-
rally marked harbor seals at Tugidak Island, Alaska. Journal of
Mammalogy 89:1201–1211.
HIBY, L., AND P. LOVELL. 1990. Computer aided matching of natural
markings: a prototype system for grey seals. Reports of the
International Whaling Commission 12:3–17.
HILLMAN, G. R., ET AL. 2003. Computer-assisted photo-identification
of individual marine vertebrates: a multi-species system. Aquatic
Mammals 29:117–123.
HUELE, R., AND H. U. DE HAES. 1998. Identification of individual
sperm whales by wavelet transform of the trailing edge of the
flukes. Marine Mammal Science 14:143–145.
KELLY, M. J. 2001. Computer-aided photograph matching in studies
using individual identification: an example from Serengeti
cheetahs. Journal of Mammalogy 82:440–449.
LATOUR, P. B. 1981. Spatial relationships and behavior of polar bears
(Ursus-maritimus Phipps) concentrated on land during the ice free
season of Hudson Bay. Canadian Journal of Zoology 59:1763–1774.
MARTIN, P., AND H. C. KRAEMER. 1987. Individual differences in behavior
and their statistical consequences. Animal Behaviour 35:1366–1375.
NICHOLS, J. D. 1992. Capture–recapture models: using marked
animals to study population dynamics. BioScience 42:94–102.
Vol. 91, No. 6
NIETFELD, M. T., M. W. BARRET, AND N. SILVY. 1994. Wildlife
marking techniques. Pp. 147–159 in Research and management
techniques for wildlife and habitats (T. A. Bookhout, ed.). Wildlife
Society, Bethesda, Maryland.
OLIVEIRA-SANTOS, L. G. R., C. A. ZUCCO, P. C. ANTUNES, AND P. G.
CRAWSHAW, JR. 2010. Is it possible to individually identify
mammals with no natural markings using camera-traps? A
controlled case-study with lowland tapirs. Mammalian Biology—
Zeitschrift für Säugetierkunde 75:375–378.
PARKS, E. K., A. E. DEROCHER, AND N. J. LUNN. 2006. Seasonal and
annual movement patterns of polar bears on the sea ice of Hudson
Bay. Canadian Journal of Zoology 84:1281–1294.
Downloaded from https://academic.oup.com/jmammal/article/91/6/1350/888329 by guest on 29 February 2024
PENNYCUICK, C. J. 1978. Identification using natural markings. Pp.
147–159 in Animal markings (B. Stonehouse, ed.). MacMillan,
London, United Kingdom.
PENNYCUICK, C. J., AND J. RUDNAI. 1970. A method of identifying
individual lions Panthera leo with an analysis of reliability of
identification. Journal of Zoology (London) 160:497–508.
PETZOLD, C. 2002. Programming Microsoft Windows with C#.
Microsoft Press, Redmond, Washington.
RAMSAY, M. A., AND I. STIRLING. 1986. Long-term effects of drugging
and handling free-ranging polar bears. Journal of Wildlife
Management 50:619–626.
REGEHR, E. V., N. J. LUNN, S. C. AMSTRUP, AND L. STIRLING. 2007.
Effects of earlier sea ice breakup on survival and population size of
polar bears in western Hudson Bay. Journal of Wildlife
Management 71:2673–2683.
VAN TIENHOVEN, A. M., J. E. DEN HARTOG, R. A. REIJNS, AND V. M.
PEDDEMORS. 2007. A computer-aided program for pattern-matching
of natural marks on the spotted raggedtooth shark Carcharias
taurus. Journal of Applied Ecology 44:273–280.
WHITEHEAD, H. 1990. Computer assisted individual identification of
sperm whale flukes. Reports of the International Whaling
Commission 12:71–77.
Submitted 30 December 2009. Accepted 15 June 2010.
Associate Editor was William F. Perrin.
APPENDIX I
The following is a summary of the steps taken by our identification
system, including image preprocessing and matching. We follow the
summary by pseudocode of the iterative adaptive thresholding
algorithm (which calls the adaptive thresholding algorithm) and the
Chamfer distance algorithm.
1. The user loads the input image, selects the side of the face (left or
right), and selects the 3 reference points: 1) front corner of the
eye, 2) notch of the nose, and 3) trailing edge of the mouth
(Fig. 1a).
2. The image is converted to grayscale using the equation 0.30 3 R
+ 0.59 3 G + 0.11 3 B, where R, G, and B are the red, green, and
blue values of a pixel.
3. The image is standardized by applying an affine geometric
transform such that the nose reference point lies 240 pixels to
the right of the eye reference point and the mouth reference
point is 128 pixels to the right and 128 pixels below the eye
(Fig. 1b).
4. The image is cropped around the whisker spot region.
December 2010 ANDERSON ET AL.—AUTOMATED IDENTIFICATION OF POLAR BEARS 1359

5. The histogram of the image is modified using histogram Adaptive thresholding (pseudocode)
matching to match the histogram of a good-quality image of Input: integer array A (grayscale image); integer r (neighbor-
polar bear whisker spots. hood radius); integer c
6. Logarithmic transformation (base e) is applied to the image. The Output: an integer array B (black-and-white image)
standard logarithmic function is p 5 c log(1 + p), where c 5 255/ set B 5 integer array with dimensions of A
log(1 + max) and max is the highest pixel value of the image. We for i 5 1, …, width of A
modified the function to account for the minimum value of an for j 5 1, …, height of A
image not always being 0, and we used p 5 c log[(1 + p)/(1 + set mean 5 mean of Ai,j neighborhood (r)
min)], where c 5 255/log[(1 + max)/(1 + min)] and min is the if Ai,j , mean 2 c then
lowest pixel value of the image. set Bi,j 5 black
7. Neighborhood averaging is applied to the image with a else
neighborhood radius of 2. set Bi,j 5 white

Downloaded from https://academic.oup.com/jmammal/article/91/6/1350/888329 by guest on 29 February 2024

8. Iterative adaptive thresholding is applied to the image with
end if
neighborhood radius of 4 and a maximum black pixel count of
end for
300 (see below for pseudocode).
end for
9. The image is cropped again around a slightly tighter whisker spot
region, and this cropped image is the final preprocessed input Chamfer distance (pseudocode)
image. Input: integer array A (black-and-white image); integer array B
10. For each preprocessed image in the reference library: (black-and-white image)
a. Let D1 be the Chamfer distance from input image to Output: real s (similarity score)
reference image. set mindistances 5 real array of size equal to the number of
b. Let D2 be the Chamfer distance from reference image to black pixels in A
input image. set n 5 1
c. Let D be the mean of D1 and D2. for i 5 1, …, width of A
d. Repeat steps a–c with one image shifted up, down, left, and for j 5 1, …, height of A
right in increments of 2 pixels for a total of 16 pixels each if Ai,j 5 black then
direction. These settings will repeat steps a–c (8 + 8 + 1)2 5 set mindistance 5 infinity
289 times. for k 5 1, …, width of B
e. Let the similarity score be the minimum of all distances for l 5 1, …, height of B
produced by step d. if Bk,l 5 black then
11. Sort similarity scores in ascending order (i.e., best to worst match) set distance 5 Euclidean distance between (i, j)
and display to the user each reference image next to its score. and (k, l)
Iterative adaptive thresholding (pseudocode) if distance , mindistance then
Input: integer array A (grayscale image); integer r (neighbor- mindistance 5 distance
hood radius); integer maxcount (maximum number of black end if
pixels desired) end if
Output: an integer array B (black-and-white image) end for
set c 5 2 end for
set count 5 infinity mindistancesn 5 mindistance
while count . maxcount n5n+1
set B 5 adaptive thresholding (A, r, c) end if
set count 5 number of black pixels in B end for
set c 5 c + 1 end for
end while set s 5 median of mindistances