This document presents a method for extracting high resolution features from internal fingerprints imaged using optical coherence tomography (OCT). The method involves fitting a curve to remove curvature from cross-sectional OCT scans of the fingertip papillary layer. Straightened cross-sections are concatenated to form a 2D internal fingerprint image, which is enhanced using filters. Minutiae features are then extracted from the image at high resolution (1300 dpi) and evaluated by matching to manually extracted features, achieving similarity scores above 0.6. The method provides a way to extract discriminating fingerprint features from internal fingerprint images without external skin limitations, but performance is still affected by fingertip motion during imaging.

1. High Resolution Feature Extraction from Optical
Coherence Tomography Acquired Internal Fingerprint
Rethabile Khutlang, Nontokozo P. Khanyile, Sisanda
Makinana
Biometrics Research Group, Identity Authentication
CSIR Modelling and Digital Science
Pretoria, Republic of South Africa
Fulufhelo V. Nelwamondo
CSIR Modelling and Digital Science
Pretoria, Republic of South Africa
fnelwamondo@csir.co.za
Abstract—Biometric fingerprint scanners scan the external
skin features onto a 2D image. The performance of the automatic
fingerprint identification system suffers if the finger skin is wet,
worn out, fake fingerprint is used et cetera. Swept source optical
coherence tomography (OCT) can be used to scan the internal
skin features, up to the depth of the papillary layer. OCT is
contactless and scans in three dimensions. The papillary contour
represents an internal fingerprint, which does not suffer external
skin problems. In this paper, we present a feature extraction
method that extracts features at high resolution from the internal
fingerprint. First curvature of an internal fingerprint cross-
section is removed by fitting a third order polynomial and
shifting each column in depth by the value of the fitted curve. A
2D image of the internal fingerprint is formed by concatenating
the individual cross-sections, averaged across the papillary
contour. The internal fingerprint image is then enhanced and
features are extracted at high resolution. We have evaluated
performance of feature extraction by matching extracted
minutiae to those extracted manually. Matching accuracy shows
that features can be extracted at high resolution from an OCT
internal fingerprint.
Keywords—Biometrics; internal fingerprint; curve fitting;
enhancement; feature extraction; optical coherence tomography
I. INTRODUCTION
Biometric identification edges out token-based and
knowledge-based identification systems on safety and
reliability because it is based on identifiers unique to
individuals. Biometric identifiers that can be used to describe
individuals should be sufficiently different for each individual
to have the requisite discrimination. They should be present on
all individuals using the identification system; and be relatively
easy to acquire. The acquired measurements should be in a
form that enables the extraction of descriptive features to be
used to identify individuals. Fingerprint biometric is the most
used physiological biometric identifier in the identification or
authentication of individuals. Of all identifiers, fingerprints
perform competitively on the factors used to assess the
suitability of any trait.
Conventional fingerprint readers capture external skin
features onto a 2-D image. External skin of palm fingers
consists of a series of ridges and furrows. The pattern of the
ridges and furrows determines the fingerprint uniqueness in
conjunction with local ridge characteristics that occur at the
ridge bifurcation and ridge ending. Fingerprint uniqueness
suffers if the ridges are worn out by heavy machinery in the
case of mineworkers for instance. Old age also negatively
affects fingerprint uniqueness.
The performance of a fingerprint recognition system is
determined by the conditions of the finger external skin first
and foremost. Performance suffers if the skin is wet, worn out,
scarred or a fake fingerprint is used [1]. Furthermore using
conventional contact-based acquisition, performance is
negatively affected by non-uniform pressure exerted by the
finger when making contact with the surface. The internal
structures of a finger – the papillary layer – can be used to
represent a fingerprint to alleviate the problems associated with
the external skin ridges and valleys. Actually the papillary
layer is the source of the fingerprint structure, and therefore
represents a blueprint of the external skin ridges and valleys.
The papillary layer, hence the internal fingerprint cannot be
destroyed by superficial cuts because it is encoded underneath
the skin surface. The fingerprint recognition system that uses
the internal fingerprint poses a challenge to fake fingerprints as
they are superficial to finger skin whereas the papillary layer is
between the dermis and the upper epidermis.
A conventional fingerprint recognition system is composed
of the sensing and acquisition, image enhancement, feature
extraction, matching and decision making components. The
same pipeline can be applied to an internal fingerprint
recognition system. Khutlang and Nelwamondo [2] used swept
source optical coherence tomography (SS-OCT) to acquire an
internal fingerprint in three dimensions contact-less. They used
novelty detection routines to segment the papillary layer, the
3D internal fingerprint was the contour of the detected
papillary region. The 3D papillary layer was detected using a
scaling-resolution edge-detection procedure in [3]. Sousedik,
Breithaupt and Busch [4] segmented a 3D internal fingerprint
using column accumulation functions. In this work, focus is
constrained to removing curvature of the 3D internal
fingerprint, enhancing the resultant 2D image and extracting
features at high resolution.
Internal fingerprint has been imaged using both time
domain OCT [5, 6], and Fourier domain OCT [7, 8]. To extract
intricate biometric details such as the distribution of sweat
pores and the pattern of the capillary bed; [9] used spectral
domain OCT, while [10] used correlation mapping OCT. Liu
and Chen [11] used Doppler OCT to image the blood vessels
down to the capillary level in the skin of a finger.
The work is sponsored by the Department of Science and Technology of
the Republic of South Africa
978-1-5090-2239-7/16/$31.00 copyright 2016 IEEE
SNPD 2016, May 30-June 1, 2016, Shanghai, China

2. Fig. 1. A 3D OCT scan (a), together with its corresponding papillary layer (b), segmented using the technique of [2].
We present a method that extracts minutiae from an internal
fingerprint image acquired contact-less using SS-OCT. First
the curvature of the 3D internal fingerprint is removed by
fitting a third order polynomial to it [9], for each cross-section.
Then the straightened cross-sections are concatenated together
to form a 2D fingerprint image. It is enhanced and minutiae are
extracted at high resolution.
II. MATERIALS AND METHODS
A swept source OCT system (OCS1300SS, Thorlabs, U.S.)
was used to capture the internal finger structure. The swept
laser optical source has a central wavelength of 1325 nm and a
spectral bandwidth of 100 nm. It has an average power output
of 10.0 mW, and an axial scan rate of 16 kHz. The system has
a maximum imaging depth of 3 mm in air. 10 mm x 10 mm
area on a fingertip was scanned using 512 axial cross-sections,
B-mode images; each 512 x 512 pixels. The resolution was
1300 dpi. Fig. 1 shows the 3D OCT scan, together with its
corresponding papillary layer, segmented using the technique
of [2].
A. Fitting a Third-Order Polynomial to Remove Curvature
The 3D contour of the papillary layer was segmented on a
cross-sectional basis, using the method of [2]. The cross-
sections were denoised using the regularised version of the
Perona and Malik PDE filter due to its stability in the presence
of speckle noise [12]. The Perona and Malik filter improves on
the heat equation (equivalent to the convolution of the signal
with Gaussians at each scale) with the signal as initial datum
by reformulating it as a nonlinear equation of the porous
medium type:
δu / dt = div (g(|∇u|)∇u), u(0) = uₒ (1)
In this equation, g is a smooth non-increasing function.
The stability brought about by [12] is due to the replacement
of the gradient |∇u| by its estimate |D * u| in the Perona and
Malik model (1).
For each filtered cross-sectional image, a third-order
polynomial was fitted to the detected contour of the papillary
layer [9]. The curvature of the fingertip was removed by
shifting each column of a cross-section in depth by the value of
the fitted curve. This straightened a B-mode scan without
affecting the undulations of the papillary contours. According
to [8], the epidermis extends to on average 0.34 mm at the
palm finger region. The straightened B-mode images were
averaged to a depth of 50 pixels, from the fitted polynomial –
Fig. 2 shows a straightened B-mode image. A 2D internal
fingerprint image was obtained by concatenating the averaged
B-mode images together, an example is shown in Fig. 3.
B. Enhancing 2D Internal Fingerpring Image
A complement of the 2D internal fingerprint image was
calculated, and the two images were enhanced. The first step in
enhancing the images was to smooth them using a Gaussian
low pass filter. Then salt and pepper noise was reduced using a
median filter. Contrast between ridges and valleys was
increased by adjusting image intensity values such that they
saturate at high and low intensities.
Finally the image and its complement were inputted to a
phase preserving denoising filter. It preserves the phase using
wavelets whose denoising results are averaged over all possible
translations of the signal to make them translation invariant
[13]. The ridges of the internal fingerprint image were eroded,
while valleys of its complement were dilated.
C. Feature Extraction
A Gabor filter tuned to the estimated local ridge orientation
and ridge frequency was applied to both the image and its
complement [14]. Regions in the filtered images that had
unreliable area or orientation were masked out. Both images
were binarized and thinned. Features were extracted from both
images, ignoring unreliable regions. Features from an image
and its complement were compared with respect to spatial
location and type, a ridge bifurcation was expected to be a
ridge ending on an opposite image. Any inconsistent features
were marked as unreliable and removed. A minutia point –
feature in an image – was represented by four attributes: row
location, column location, whether it is a bifurcation or ridge
ending, and the minutia angle. With the minutia point used as

3. Fig. 2. A curve fitted to the papillary contour and the resulting straightened
image.
reference, the angle is measured clockwise from the horizontal
axis to:
• A line parallel to the ridge that just ended at the
minutia location
• A line dividing the smallest angle formed by two
ridge that had met at a ridge bifurcation
D. Evaluation of Extracted Features
Features were manually extracted from the internal
fingerprint images. The manually extracted features were used
as a gold standard in assessing the performance of the
automated feature extractor developed. A similarity score when
matching manual features to automated ones was used as a
measure of how well the automated feature extraction
performed; using the similarity score formulated by [15]. The
similarity score is between 0 and 2, for a perfect non-match
and a perfect match respectively. A similarity score above 0.4
is usually regarded as a match. It was evaluated using the
following function for two minutiae point representations:
=
2
( + )
×
1
+
1
(2)
where is the number of mated minutiae points. and
is the number of the first input minutiae points and the number
of those minutiae points that lie within the region of overlap
between the input minutiae points after alignment,
respectively; and similarly for and .
III. RESULTS
A third order polynomial was fitted to the detected contour
of the papillary layer. Curvature was removed by shifting each
column in depth by the value of the fitted curve. Fig. 2 shows
both the curve fitted using the papillary contour and the
resulting straightened image. The 2D internal fingerprint image
was obtained by concatenating the straightened B-mode
images, which were averaged to a depth of 50 pixels –
determined empirically.
An internal fingerprint image was put through a Gaussian
low-pass filter with a correlation kernel of size 7x7 pixels and a
standard deviation sigma of 0.7. The median filter inputted the
smoothed image; each output pixel contained the median value
in a 7x7 neighbourhood around that pixel. It was empirically
discovered that saturating 0.3% of image intensity values at
high intensities and 0.8% at low intensities yields the best
contrast adjustment. Fig. 3 shows a contrast adjusted internal
fingerprint image. The phase preserving filter rejected noise
above two standard deviations. It used 6 filter scales to cover
low frequencies. Both erosion and dilation used a 45° ,
measured counter-clockwise from the horizontal axis, line of
length three pixels as a structuring element.
Fig. 3. A contrast adjusted internal fingerprint image.

4. Fig. 4. Images (a) and (c) have minutiae extracted using the workflow presented overlaid; images (b) and (d) are equivalent images with manually extracted
minutiae overlaid. Yellow lines show points that were matched; convex hull points are also shown in green.
TABLE I. PERFORMANCE EVALUATION OF FEATURE EXTRACTION FOR
IMAGES IN FIG. 4.
Imag
es Similarity
score
Matched
points
Boundary
points set
A
Boundary
points set
B
Overlappi
ng region
points
A&B 0.7692 5 9 6 5
C&D 0.6250 5 7 8 5
Extracted features were overlaid on a 2D internal
fingerprint image. Blue cycle annotates a bifurcation, green
cycle a ridge ending. Fig. 4 shows matched minutiae, and
Table I gives matching parameters for the two images.
IV. DISCUSSION
A third order polynomial fitted well to the papillary
contour. The fit of a curve to the contour is affected by the
accuracy of the papillary contour detection. The fitted curve
averages the ridge and valley undulations of the internal
fingerprint to cut through their midpoint. Significant
discrepancies in the detected papillary contour will shift the
curve away from an internal fingerprint and the average taken
will not be around the upper edge of the papillary layer.
Even though the individual B-mode images comprising a
3D OCT scan were denoised using a regularized version of the
Fig. 5. Effects of fingertip motion on internal fingerprint feature extraction. The integrity of ridges and valleys is affected by the fingertip motion. (a) shows a 3D
OCT scan of a fingertip. (b) shows the internal fingerprint image obtained by straightening the papillary contour and averaging to a depth of 50 pixels, the image
has been adjusted for contrast. (c) shows the binarised image, after applying a phase preserving filter. The green arrow shows an image artefact caused by motion.

5. Perona and Malik partial differential equations filter, as part of
the papillary contour detection [2]; it was observed that the
concatenated internal fingerprint image was noisy and contrast
was poor. It was discovered that the combination of a low pass
filter and a median one minimises the noise. And the
percentages of pixels to be saturated to high and low intensities
to pronounce the ridge valley separation were determined
empirically. The phase preserving filter further enhanced the
images.
The regions in the Gabor filtered image that had unreliable
orientation were masked out. That resulted in fewer minutiae
extracted from areas that did not have clear structure of ridges
and valleys. Distance between each pair of ridge endings and
their angles were calculated to determine if the pair could be
from a broken ridge; if so they are removed from the list of
extracted features. Using the complement of the internal
fingerprint image reduced the amount of spurious minutiae that
would otherwise have been extracted.
Minutiae are typically extracted from 500 dpi fingerprint
images that represent external fingertip skin undulations. In
this work, minutiae are extracted at high resolution, 1300 dpi
from a fingerprint acquired within fingertip skin, at the
papillary layer region. These internal fingerprint images will be
reduced to 500 dpi so that extracted minutiae can be compared
to those extracted from external skin fingerprint images using
conventional fingerprint feature extraction methods.
The biggest source of error in the feature extraction is the
movement of a fingertip during image acquisition. If the hand
moved while OCT was scanning a fingertip, it resulted in
image artefacts. Fig. 5 shows the effect of such movement. The
acquired image has discontinuities due to such movement.
Since papillary contour detection is based on the corneum
stratum, the discontinuities results in internal fingerprint image
artefacts. Improvements in the papillary contour detection
method [2] will improve the proposed high resolution feature
extraction pipeline. Such improvements will yield a smoother
papillary contour. The proposed 2D internal fingerprint
segmentation method struggled towards the edges of an image,
example images in Fig. 4. This is mainly due to intensity
depth-dependent roll off. One method of overcoming the roll-
off problem is to use a glass slide during scanning [3].
The similarity scores of two images are reported. The first
step in the development of the proposed method will be to
create a significant dataset of fingertip acquisitions in order to
conduct a more thorough performance evaluation of feature
extraction. The OCT scanner used to acquire an internal
fingerprint cannot cover an entire area of a typical fingertip. A
10 mm X 10 mm area was acquired. This has resulted in low
similarity scores as few minutiae could be extracted from the
area imaged. An improvement of the area that the OCT
machine covers will result in high similarity score when
matching automatically extracted minutiae to manual ones as
more minutiae will be extracted from a bigger scanned area of
a fingertip. Improvement of acquisition speed will also result in
high similarity scores as finger movement will be minimised
during acquisition.
V. CONCLUSION
The proposed workflow automatically fits a third order
polynomial curve to the papillary contour for each B-mode
image. The curvature of the fingertip is removed by shifting
each column of a B-mode image in depth by the value of the
fitted curve. This straightens the papillary contour. The
straightened papillary is averaged to a depth of 50 pixels, and
the averaged cross-sections are concatenated together to form a
2D image of the internal fingerprint. The 2D image is enhanced
to increase contrast between ridges and valleys; then minutiae
are extracted at high resolution. The similarity score obtained
when matching extracted minutiae to those manually extracted
confirm that the extracted features are usefully for matching.
REFERENCES
[1] A. Jain, A. Ross, and S. Prabhakar, “An introduction to biometric
recognition,” IEEE Trans. Cir. Sys. Video Tech., vol. 14, pp. 4-20, Jan.
2004.
[2] R. Khutlang and F. Nelwamondo, "Novelty detection-based internal
fingerprint segmentation in optical coherence tomography
images", Computing and Networking (CANDAR), Second Inter.
Symposium on. IEEE, pp 556 – 559, Dec. 2014.
[3] L. Darlow and J. Connan, “Efficient internal and surface fingerprint
extraction and blending uisng optical coherence tomography”, Applied
Optics, vol. 54(31), pp. 9258 – 9268, Nov. 2015.
[4] C. Sousedik, R. Breithaupt, and C. Busch, “Volumetric fingerprint data
analysis using optical coherence tomography,” Inter. Conf. BIOSIG,
pp.1-6, Sept. 2013
[5] Y. Cheng and K. Larin, “Artificial fingerprint recognition by using
optical coherence tomography with autocorrelation analysis,” Appl.
Opt., vol. 45, pp. 9238-9245, Dec. 2006.
[6] Y. Cheng and K. Larin, “In vivo two- and three-dimensional imaging of
artificial and real fingerprints with optical coherence tomography,”
IEEE Photonics Tech. Letters, vol. 19, pp. 1634-1636, Oct. 2007.
[7] S. Dubey, D. Mehta, A. Anand, and C. Shakher, “Simultaneous
topography and tomography of latent fingerprint using full-field swept-
source optical coherence tomography,” J. Opt. A: Pure Appl. Opt., vol.
10, pp. 1-8, Jan 2008.
[8] A. Bossen, R. Lehmann, and C. Meier, “Internal fingerprint
identification with optical coherence tomography,” IEEE Photonics
Tech. Letters, vol. 22, pp. 507-509, April 2010.
[9] M. Liu and T. Buma, “Biometric mapping of fingertip eccrine glands
with optical coherence tomography,” IEEE Phot. Tech. Lett., vol. 22,
pp. 1677-1679, Nov. 2010.
[10] A. Zam, R. Dsouza, H. Subhash, M. O’Connell, J. Enfield, K. Larin, and
M. Leahy, “Feasibility of correlation mapping optical coherence
tomography (cmOCT) for anti-spoof sub-surface fingerprinting,” J.
Biopho., pp. 1-5, April 2013.
[11] G. Liu and Z. Chen, “Capturing the vital vascular fingerprint with
optical coherence tomography”, Applied Optics, vol. 52(22), pp.5473 –
5477, Aug. 2013.
[12] F. Catte, P. Lions, J Morel, and T. Coll, “Image selective smoothing and
edge detection by nonlinear diffusion,” SIAM J. Num. Anal., vol. 29,
pp. 182-193, Feb. 1992.
[13] P. Kovesi, “Phase preserving denoising of images”, The Australian Patt.
Recog. Soc. Conf., pp. 212 – 217, Dec. 1999
[14] L. Hong, Y. Wan and A.Jain, “Fingerprint image enhancement:
algorithm and performance evaluation”, IEEE Trans. on Patt. Ana. and
Machine Int., vol. 20(8), pp. 777 – 789, Aug. 1998.
[15] N. Khanyile, A. Kock, and M. Mathekga. "Similarity score computation
for minutiae-based fingerprint recognition", Biometrics (IJCB), 2014
IEEE International Joint Conference on. IEEE, pp. 1 – 8, Sept. 2014.