INTRODUCTION:

Automated identification of humans using their unique anatomical characteristics has been
increasingly investigated for their applications in human surveillance and image forensics.
Emerging national ID programs that require accurate, online and large scale identification
automated personal identification have posed new challenges for the biometrics technologies.
The unique identification is one such ambitious project that aims to identify ~1.2 billion
population using ten fingerprints and two iris images. Selection of biometrics modalities in such
large scale identification problems is not only limited by the individuality of the modality but
also by the user-convenience in acquiring the respective modality. In this context, the finger-vein
and finger knuckle images can be simultaneously acquired while acquiring the fingerprint images
and with no additional inconvenience to the users. Simultaneous acquisition of finger-vein
images can however require some alterations in the existing (slap) fingerprint devices, largely
due to the near infrared based intrusive imaging requirements for finger-vein imaging. However,
the finger knuckle images can be simultaneously acquired with the addition of an external
imaging camera that simultaneously acquires finger dorsal images and synchronizes the
acquisition with external software. Therefore it is important to ascertain the nature of information
that can be extracted from the finger dorsal images. This paper focuses on this problem and
investigates the possibility of using minor finger knuckle patterns for the biometric
identification.

Accurate identification of finger knuckle patterns can be beneficial for several

applications involving forensic and covert identification of suspects. There are several classes of
forensic images in which the finger knuckle patterns are the only piece of evidence available to
identify the suspects. Figure 1 shows some examples of the photographs in which the finger
knuckle pattern is the only or major source of information available to scientifically ascertain the
identity of individuals. Therefore the matching of finger knuckle patterns can help to identify the
suspects and ascertain supportive scientific evidence from the photographs, especially in cases
when no information regarding fingerprint or face is present in the available photographs.
 (a) (b)

(C) (d)

Figure 1: Sample photographs with knuckle patterns during (a) kidnapping,

(b) covert video and surveillance, (c)-(d) emotions for surveillance and forensic applications.

The legal issues relating to the reliability of finger knuckle image patterns will largely be judged
in the courtrooms. Therefore any new biometric to be introduced for the human identification
should also meet the requirements stipulated by courts to be deemed admissible. Such
requirements can vary among different courtrooms but often require reliable and repeatable
measurements. It is therefore important that any new/potential biometric evidence to be
admissible by court, in addition to their uniqueness, their stability over a reasonable time period
should also be established. A preliminary study presented in this paper is motivated to check the
veracity of the questions and assertions (we received during peer reviews on earlier papers [3],
[7]) that the stability of finger knuckle patterns, especially for forensic and law-enforcement has
never been explored/established.
Finger Knuckle
A normal human hand has four fingers each of which has 3 bone segments and 3 joints. The
thumb has 2 bone segments and 2 joints. These segments are known as phalanges (plural of
phalanx) and are shown in figure 2 from a typical finger dorsal image. While in some humans the
major finger knuckle pattern can be occluded by hair, the minor finger knuckle patterns do not
appear to suffer from such problem. There are several forensic images when only minor finger
knuckle patterns/portions are visible/available for any possible identification. In addition, the
matching results from the minor finger knuckle matching can also be employed to improve the
reliability and accuracy of conventional/emerging major finger knuckle based biometric
identification.

Figure 2: Image sample from a typical finger dorsal surface image identifying the major and
minor knuckle pattern regions with respect to the MCP/DIP joints.

Processing Steps:

Image Segmentation and Normalization

Accurate personal identification using minor finger knuckle patterns will require accurate
segmentation of region of interest images. The segmentation approach should be able to generate
normalized and fixed size region of interest images from the finger dorsal images of subjects
under varying age group. In absence of any fixation pegs or the finger docking frame, the
acquired finger dorsal images illustrate fingers with varying poses, locations and scale changes.
In addition, the varying length of fingers, finger-widths, finger-nails, skin pigmentation and
location of distal interphalangeal points, poses severe challenges to exploit any anatomical
characteristics of fingers for robust minor finger knuckle segmentation. Figure 3 illustrates
simplified block diagram for the finger knuckle segmentation strategy attempted in this work to
segment fixed size minor finger knuckle images.

Figure 3: Simplified block diagram illustrating key steps in the automated segmentation
of minor finger knuckle images from the finger dorsal images.
Each of the acquired images is firstly subjected to binarization using Otsus thresholding. The
resulting images are cleaned (denoised) by automatically removing the isolated regions/pixels (<
100 pixels) so that the longest object representing finger is only retained. The binarized finger
shape is used to estimate the location of finger-tip from the convex hull of the images. The
location of finger-tip is utilized to eliminate the background image above the finger tip. The
orientation of fingers is then estimated from this binarized image using the methods of moment,
similar to as also employed in [5]. This step is followed by the coarse segmentation which
segments a small portion of acquired finger images that can include minor finger knuckle region
while excluding major knuckle region and major part of finger nail. Such segmentation strategy
requires some crude assumptions for the maximum ratio of nail length to the finger length and
assumption that the major finger knuckle region is located somewhere in the middle of the
acquired finger dorsal image. The resulting coarsely segmented image is further subjected to nail
check and removal steps which consist of segmenting the image and locating the bonding box
region for smaller parts and removing them. The width of the resulting image is computed and
used to estimate the scale factor for the scale normalization. The edge detection of resulting
image is used to locate the center of minor finger knuckle image. This is achieved by estimating
the location of the centroid for the resulting edge detected image and segmenting a fixed size

region (160 180 pixels) that represents minor finger knuckle region for the finger dorsal image.

Image Enhancement

The finger dorsal surface is 3D curved surface and such curves can result in uneven illumination
reflections and shadows. Therefore the segmented minor finger knuckle images often have low
contrast and illumination variations. The enhancement steps are essentially required to normalize
such illumination variations. The illumination normalization approach used in this work is same
as also used in [7]. This approach firstly estimates the average background illumination in the 16

16 pixels sub-blocks of the segmented knuckle images. The estimated illumination is then
subtracted from the original knuckle image to remove the uneven illuminations. The resulting
image is then subjected to the histogram equalization operation which generates enhanced minor
finger knuckle image for the feature extraction stage. Figures 4 (d1)-(d5) shows image samples
after the image enhancement operations.

Feature Extraction and Matching

The finger knuckle images after enhancement typically represent some random texture pattern
which appears to be quite unique in different fingers. Therefore a variety of spatial and spectral
domain feature extraction strategies can be pursued to ascertain the matching accuracy from the
minor finger knuckle images. The experimental results in this paper have employed local binary
patterns, improved local binary patterns, band limited phase only correlation and 1D log-Gabor
filter based matchers for the performance evaluation.
 Related Works

The use of finger knuckle images for the biometrics identification has generated increasingly
interest in the literature. Woodard and Flynn [1] successfully demonstrated the use of 3D finger
dorsal images for personal identification. This work essentially exploits local curvature patterns
on the 3D finger surface and quantifies them into various shape indexes for the matching.
Reference [3] details an online system using the hand dorsal surface images which can
simultaneously exploit the finger knuckle patterns from the multiple fingers and their
geometrical shape characteristics. There are several publications which have exploited the
effectiveness of finger knuckle patterns using contactless imaging [1], [3], [7], [9] and contact
based or constrained imaging [8], [11] [17]. These references in the literature have however
exploited major finger knuckle images which capture patterns formed on the finger dorsal
surface joining proximal phalanx and middle phalanx bones. In the best of our knowledge, there
are no known efforts to exploit minor finger knuckle patterns (figure 1) which are formed on the
finger dorsal surface joining distal phalanx and middle phalanx bones.

The unidirectional bending of fingers is primarily responsible for generating skin pattern
alterations on the finger dorsal surface joining the four phalanx bones. The minor finger knuckle
patterns are formed on the surface joining distal phalanx and middle phalanx bones and can also
be quite distinctive for biometrics identification. This paper [23] has attempted to examine
biometric identification capability for humans using such minor finger knuckle images and
develops effective algorithms for the automated segmentation of region of interest, image
normalization, enhancement and robust matching to accommodate inherent image variations.
Key contributions from this paper can be summarized as in the following:

(a) This paper investigates on the possibility of using minor finger knuckle patterns for human
identification. A completely automated scheme to simultaneously segment minor and major
finger knuckle images from contactless finger dorsal images is developed. Combination of
simultaneously acquired minor finger knuckle pattern and major finger knuckle pattern images
can achieve significant improvement in performance, which is not possible by using major finger
knuckle images alone as in the literature.
 (b) Lack of any systematic study to ascertain stability of knuckle patterns raises several
questions on the possible use of finger knuckle patterns in image forensics for law
enforcement and civilian applications. This paper therefore also presents a/first study
(also publicly providing such images for further investigation) to ascertain the stability of
knuckle pattern in finger dorsal images acquired over the interval of over 6 years. Such
study is especially important for forensic analysis of those images in which finger
knuckle is the only piece of evidence available to identify the suspects.

(c) This paper provides first publicly available database on minor knuckle images (also
major finger knuckle images), from 503 different subjects. In the best of our knowledge
this will be largest subjects database available to-date in public domain and help to
advance further research efforts in this area.

The use of finger knuckle print as biometric identifier has generated increasingly interest in the
literature. Woodard and Flynn successfully demonstrated the use of 3D finger dorsal images for
personal identification. This work essentially exploits local curvature patterns on the 3D finger
surface and quantifies them into various shape indexes for the matching. A. Kumar, senior
member, IEEE, [1] This paper has successfully investigated the likelihood of employing minor
finger knuckle images for the identification. The coarse-to-fine segmentation strategy developed
in this paper has been quite self-made because it has been able to achieve higher matching
accuracy. The experimental results illustrated during this paper, on the info of 503 subjects, can
achieve promising performance (EER of 6.29% and 12.6% under two protocols) from contactless
minor finger knuckle images. The experimental results according during this paper conjointly
recommend that the synchronous use of major and minor finger knuckle images will help to
considerably improve the performance which will not be attainable by victimization either minor
or major finger knuckle images alone. Kumar and C. Ravikanth [2] details an online system
using the hand dorsal surface images which can simultaneously exploit the finger knuckle
patterns from the multiple fingers and also their geometrical shape characteristics. Better results
were found out with different fusion of basic techniques as Principal Component Analysis
(PCA), Linear Discriminate Analysis (LDA) and Independent Component Analysis (ICA).
This method overcome problem due to finger rings and black background. Previous method used
black background problem of using dark background is that the finger surface can be dark, so
they used white background which requires simple pre-processing as compared. The
disadvantage of this method is the speed of working is less as it uses scanner for imaging. A.
Kumar, [3] in this paper they work on a new approach to enhance the performance of finger vein
identification System. The projected system at the same time acquires the finger vein and low
resolution finger print images and combines these two evidences employing a novel score level
combination strategy. The utility of low resolution fingerprint pictures non inheritable from a
digital camera is examined to establish the matching performance from such pictures. They
developed and investigated two new score level combos, i.e., holistic and nonlinear fusion, and
relatively assess them with additional well-liked score level fusion approaches to establish their
effectiveness in their projected system. They presented the information of 6,264 pictures from
156 subjects illustrate vital improvement within the performance, each from the authentication
and recognition experiments. S. Aoyama, K. Ito, and T. Aoki, [5] in this paper they work on FKP
recognition based on Band Limited Phase Only Correlation (BLPOC). POC is an imaging
matching technique using the phase components in 2D DCT of given images. BLPOC is a
modified version of POC which is dedicated to evaluate similarity between images, in order to
handle the non-linear deformation of FKP images. Most of POC based biometric recognition
algorithms cannot handle the nonlinear deformation of pictures, since the part info obtained from
the whole image is used. So as to handle the nonlinear deformation of FKP pictures, the planned
algorithmic program employs native block matching exploitation BLPOC, since the nonlinear
deformation is around diagrammatically by the minute translational displacement between native
image blocks. The Region of Interest (ROI) is extracted from the FKP image within the pre-
processing. The translational displacement between the 2 ROI pictures is calculable victimization
BLPOC and also the two images area unit aligned consistent with the calculable displacement.
Then, the common region of the two images is extracted. If the world magnitude relation of the
common region between the ROI pictures is below the edge, the BLPOC operate between the
ROI images is calculated. Otherwise, the BLPOC operate between the common regions is
calculated. Finally, the average BLPOC function is calculated from all the corresponding point
pairs regardless of the reliability of correspondence. Kam Yeun Cheng, Ajay Kumar,[12]This
paper details the development of a smart phone based online system to automatically identify a
person by using their finger knuckle image. The key objective is to exploit user-friendly
biometric, with least privacy concern, to enhance security of the data in smart phone. The
developed system application can operate on any smart phones which use android OS and
contain an embedded camera (with at least 2MP resolution). This paper uses 1D log-Gabor filter
to extract the finger knuckle templates which are matched using Hamming distance. The
limitation with this system is that for accurate finger knuckle detection, the acquire image
background should be largely uniform as the background noise would influence the auto finger
detect capability. This paper has also developed a new smart phone based finger knuckle image
database of 561 finger knuckle images of 187 different fingers from 109 users, in real imaging
environment.
EXISTING WORK:

Finger or palm dorsal surface is inherently revealed while presenting (slap) fingerprints during
border crossings or during day-to-day activities like driving, holding arms, signing documents or
playing sports. Finger knuckle patterns are believed to be correlated with the anatomy of fingers
that involve complex interaction of finger bones, tissues, and skin which can be uniquely identify
the individuals. This paper investigates the possibility of using lowest finger knuckle patterns
formed on joints between the metacarpal and proximal phalanx bones for the automated personal
identification. We automatically segment such region of interest from the palm dorsal images
and normalize/enhance them to accommodate illumination, scale and pose variations resulting
from the contactless imaging. The normalized knuckle images are investigated for the matching
performance using several spatial and spectral domain approaches. We use database of 501
different subjects acquired from the contactless hand imaging to ascertain the performance. This
paper also evaluate the possibility of using palm dorsal surface regions, along with their
combination with minor knuckle patterns, and provides palm dorsal image database from 712
different subjects for the performance evaluation. The experimental results presented in this
paper are very encouraging and demonstrates the potential of such unexplored minor finger
knuckle patterns for the biometrics applications.

FINGER KNUCKLE IMAGING AND REGION OF INTEREST SEGMENTATION

The key region of interest in this investigation represents region of image between metacarpal
and the proximal phalanx from the finger dorsal surface. This region of interest is automatically
segmented from the hand dorsal images employed for this work. The hand dorsal surface images
from the right hands of volunteers were acquired using contactless hand imaging. The imaging
setup is similar to as the one employed in reference [2]. However in addition to the indoor
illumination, the majority of images in the developed database were acquired under outdoor
(ambient) illumination. Acquired images are firstly subjected to histogram equalization,
binarization (Otsus method followed by removal of isolated noisy pixels) and used to generate
hand contour images as show n in figure 3. The key objective in this work is to evaluate the
uniqueness and stability of second minor finger knuckle patterns which are formed on the skin
surface above the middle phalanx and proximal bone joints of finger dorsal surface. Therefore
the automated localization of key points from the hand contour images, similar to as detailed in
reference [2], is performed and utilized in this work (not described here as more details can be
referred from [2]). Majority of images employed in this work were acquired under outdoor
environment u sing hand held camera. The distance between the hand held digital camera and
hand dorsal surface in contactless imaging is not fixed. Such variations in the distances often
generate hand dorsal images with varying scale and therefore scale normalization of the acquired
images is performed. This is achieved by normalizing the acquired images to a fixed scale. The
scale factor for such normalization is computed from the ratio of distance between the two finger
valleys (v1 and v3 in figure 3) and a fixed distance computed from the average of such finger
valley distances from sample images (fixed to 325 in all our experiments).

Figure 3: Automated segmentation of region of interest (second minor finger knuckle); (a) acquired image sample
under outdoor illumination, (b) recovered hand contour image (enhanced for easy visualization), and (c) localization
of key points. The red squares in figure (d) illustrate the localized second minor finger knuckle region and the
corresponding segmented knuckle images are shown in (e).

The key points corresponding to (four) finger tips and mid -point of two base points for each of
the finger are used to further localize the second minor finger knuckle region of interest. The line
joining finger-tip (top point) and the base point (mid -point of v1 and v2) is fu rther extended by
an amount one third of finger length (or by the distance between M23 and M14 in figure 3). This
extended point (R in figure 3) is used as the center point of a 100 100 pixel square region
which is automatically segmented as the second minor finger knuckle image. Figure 4 illustrate
another two image samples from our d atabase and their automatically localized second minor
knuckle regions are also highlighted as the red color boxes.

Figure 4: Image samples from two volunteers acquired under outdoor (a) and in indoor (b) environment. The red
boxes 100illustrate automatically localized and segmented 100 pixels ROI to ascertain their possible usage as a
biometric trait.

Acquired images are firstly subjected to histogram equalization, binarization (Otsus method
followed by removal of isolated noisy pixels) and used to generate hand contour images as show
n in figure 3. The key objective in this work is to evaluate the uniqueness and stability of second
minor finger knuckle patterns which are formed on the skin surface above the middle phalanx
and proximal bone joints of finger dorsal surface. Therefore the automated localization of key
points from the hand contour images, similar to as detailed in reference [2], is performed and
utilized in this work (not described here as more details can be referred from [2]). Majority of
images employed in this work were acquired under outdoor environment u sing hand held
camera. The distance between the hand held digital camera and hand dorsal surface in
contactless imaging is not fixed. Such variations in the distances often generate hand dorsal
images with varying scale and therefore scale normalization of the acquired images is performed.
This is achieved by normalizing the acquired images to a fixed scale. The scale factor for such
normalization is computed from the ratio of distance between the two finger valleys (v1 and v3
in figure 3) and a fixed distance computed from the average of such finger valley distances from
sample images (fixed to 325 in all our experiments). The key points corresponding to (four)
finger tips and mid -point of two base points for each of the finger are used to further localize the
second minor finger knuckle region of interest. The line joining finger-tip (top point) and the
base point (mid -point of v1 and v2) is fu rther extended by an amount one third of finger length
(or by the distance between M23 and M14 in figure 3). This extended point (R in figure 3) is
used as the center point of a 100 100 pixel square region which is automatically segmented as
the second minor finger knuckle image. Figure 4 illustrate another two image samples from our
database and their automatically localized second minor knuckle regions are also highlighted as
the red color boxes.

Image Normalization
The finger dorsal images are acquired from the curved 3D knuckle surface and such curves can
result in uneven illumination reflections and shadows. Therefore the segmented minor finger
knuckle images often have low contrast and illumination variations. The image enhancement
steps are essentially required to normalize such illumin ation variations in the ROI images. The
illumination normalization approach used in this work is same as also used in [7]. This approach
firstly estimates the average 16 pixels sub-blocksbackground illumination in the 16 of the
segmented knuckle images. The estimated illum ination is then subtracted from the original
knuckle image to suppress the influence of uneven illuminations. The resulting image is then
subjected to the histogram equalization operation which generates enhanced finger knuckle
image for the next or feature extraction stage. Figures 4 (b1)-(b5) and 4(c1)-(c5) illustrate second
minor finger image samples before and after the image enhancement operations respectively.
These images illustrate that employed enhancement approach has been quite effective for
enhancing the knuckle creases and curves from the automatically segmented second minor
knuckle images.
FEATURE EXTRACTION AND MATCHING
The second minor finger knuckle images after the image enhancement illustrate randomly
textured patterns which appears to be quite unique in images from different fingers/ subjects.
Such patterns typically consist of creases, lines, and w rinkles of varying thickness, which
alsovaries with the forward movement of respective fingers. A new approach to match such
second minor knuckle images is investigated and detailed in the next section. We also
comparatively evaluated several matching strategies which have been shown [6], [16]-[19] to be
effective in matching palm or major finger knuckle patterns in the literature. These are also
briefly described in following.

Figure 5: Automatically segmented lower or second minor knuckle images corresponding to middle finger images of
five different subjects in (a 1 -a5 ) respectively. Corresponding grey-level images in (b1 -b5 ), enhanced images (c1 -
c5 ), KnuckleCode or RLOC representation (d 1 -d5 ), Ordinal representation for 30 deg (e1 -e5 ), and spectral
representation using BLPOC in (f1 -f 5 ).

Local Feature Descriptor

The second minor finger knuckle images typically illu strate texture-like details which are
random but appears to be quite unique for each of the fingers. These texture patterns can be more
effectively and efficiently matched from their spatial feature similarity in local regions.
Therefore a new spatial-domain approach [27] to match the knuckle patterns was investigated in
this work. A spatial filter is employed to perform convolution with the knuckle images and the
quantization of resulting filter response generates respective features corresponding to chosen
quantization levels. The two-level quantization can provide compact or smallest size template
representation and also generate fast similarity scores using computationally simpler Hamming
distances. The features generated from such spatial filtering are expected to encode features
which are localized and represents the neighborhood pixel information. The size of this spatial
filter defines the scope of this neighborhood. The larger (smaller) filter size is expected to encode
more global (local) information and vice versa. Let the shape of this spatial filter be a square to
ensure symmetry, then the simplified filter (, ) can be constructed as follows:

where , is index in the filter, , [, ], and defines the scope of neighborhood. The
filtered response from a point on knuckle surface is essentially the sum of difference between the
point and each of its neighboring points. If the local grey level continuity is to be ensured, like in
knuckle or natural images, the difference between grey levels from two points which are close to
each other should almost be zero. Therefore the sum of them should also be very small and the
binary feature (, ) for the point (, ) in knuckle image can be computed as follows:

where represents pixel-w ise convolution operation. When the encoded feature from (2) is one,
the local knuckle shape is expected to concave else the described shape is more likely to be
convex. It is generally believed that the gradient of descriptors can offer more powerful
capability to describe the feature. Therefore the gradient version of filter (, ) in (1) is expected
to generate more powerful feature descriptor. We partition the neighbors of a given point (, )
into two subsets whose spatial extent or size is the same. For all the points or image pixels in the
two subsets, we subtract their value from (, ). After computing the sum of points values for
each subset, the gradient is defined as the difference between these two respective/ sum results.
Let these two subsets be represented as 1 and 2 , then the gradient filter (, ) obtained from
(, ) can be defined as follows:

There can be several possibilities to explore the spatial extent of partitions 1 and 2 . In this
work we only considered simplified partitions to define (, ) as follows:

(4)
where , is index in the filter with , [, ] and () is the absolute operation. The matching
distance between the two feature matrix, i.e., one obtained from the gallery (G) and the probe (E), is
computed as follows:

where (, ) is the Boolean XOR operator that computes Hamming distance between two
binarized template G and E while m and n denotes the spatial size of these templates. In order to
accommodate possible translations between the knuckle surface, we translate the binarized
templates of the probe in horizontal q and vertical direction p and perform multiple matches.
These translations are performed in two steps for four directions (left, right, up, down) in steps of
two pixels. The best or smallest of Hamming distance among these translations is used as final
similarity score (5) between the two matched templates.

Band Limited Phase Only Correlation

Another possible approach to match these knuckle patterns is to compute their similarity or
correlation from their spectral representations. Such an approach has shown to offer more
accurate results [16]-[17] and was therefore attempted to match second minor knuckle images.
This approach only uses phase information recovered from the 2D discrete Fourier transform
(DFT) of the segmented knuckle images and therefore least sensitive to the translation/ rotational
changes. In order to minimize the influence noise, only a band of frequency in the DFT
representation is employed for the matching. This approach detailed in [17] is briefly
summarized in the following. Let 2D DFT of two P Q pixels normalized knuckle images, say
S1 (x, y) and S2 (x, y), be respectively represented as F1 (k1 , k2 ) and F2 (k1 , k2 ).

Local Radon Transform

Matching knuckle images using Local Radon Transform (LRT), referred to as RLOC [19] or
KnuckleCode representation in [6], has shown to be quite effective in accurately matching the
major knuckle patterns and was therefore also evaluated in this work. This approach effectively
encodes the local orientation of curved lines and knuckle creases into one of the dominant
orientations which is represented using a three bit binary code and such binarized templates are
matched using the Hamming distance. The details of this approach can be found in reference [6],
[13] or [19]. The key advantage of this approach is that it produces smaller template size [13]
and is computationally efficient in generating templates than BLPOC, and also the ordinal
representation approach considered for the performance comparison. Another approach referred
as (fast) CompCode [13], w hich is quite similar in encoding local knuckle crease orientations bu
t using even Gabor filters, w as also evaluated for the comparative performance.
Ordinal Representation
The ordinal measures typically compute measurements that are based on the relative distances
[20]. The ordinal representation of textured like surface such as palm and iris has show n to offer
promising results and therefore this method w as also evaluated to ascertain its accuracy in
matching the knuckle images. Similar to as detailed in reference [18], two orthogonal Gaussian
filters oriented at 0, 30 and 60 degrees are utilized to generate binarized feature templates. The
Hamming distance between resulting feature templates is employed to generate match scores
between the knuckle images matches.

EXPRIMENTS AND RESULTS

The experiments were performed in several phases to ascertain the usefulness of second minor
finger knuckle patterns for the biometric authentication. We acquired hand dorsal images from
501 different subjects using contactless hand imaging. The images were acquired from the right
hand of the volunteers under indoor or outdoor environment. All the acquired images were used
to automatically segment 100 100 pixel ROI corresponding to second minor finger knuckle
region as described in section 2. These images were enhanced and subjected to the feature
extraction using three matchers as discussed in section 3. The band limiting threshold for
BLPOC was fixed to 0.6 for all the experiments. The line width and length of one and seven
pixel was respectively fixed for all experiments in RLOC (the filter y = 1) ofx = 3, 11). Two
Gaussian filters (size was 11 size 11 11 pixels were utilized to compute ordinal
representations of the segmented knuckle for the matching. Figure 5 illustrates some sample
knuckle images and corresponding template images for respective enhanced knuckle images. The
templates corresponding to the computationally simpler method detailed in section 3.1 are shown
in figure 6.
Figure 6: Typical samples of templates generated from six different subjects second minor knuckle using
local features

The first sets of experiments were performed to ascertain the suitability of matcher and its
performance for the second minor finger knuckle patterns. We utilized five images from each of
the four fingers, for every subject, and employed relatively challenging protocol (all-to-all
instead of leaveone-out) for matching second minor knuckle images. Therefore 5010 (501 10)
genuine scores and 1252500 5500) impostor scores were computed. The matching(501
performance was evaluated using the receiver operating characteristic (ROC). The ROC
corresponding to the index, little, middle, and ring finger is shown in figure 7(a), (b), (c) and (d)
respectively. It can be ascertained from these ROCs that the local feature based matcher
evaluated in this work performs the best among five matchers considered for the performance
comparison. The equal error rate (EER) from different matchers is also presented in table 1. In
terms of EER, the performance of this approach can be considered to be similar to the BLPOC
although BLPOC marginally performs better in some cases. It should however be noted that
ROC rather than EER is widely considered as the reliable criterion for performance comparison.
The complexity of BLPOC that requires computation of 2DFFT is also significantly high.
Therefore we employed only the outperforming matcher (figure 7) using local features for other
further experiments reported in this paper.
Summary:
This paper has investigated the possibility of using second minor finger knuckle image for the
personal identification. The approach described in this paper is completely automated and uses
contactless imaging which is expected to produce/ accommodate large variations in images. The
experimental results on the image database from over 500 subjects presented in this paper
suggest great potential for the second minor knuckle patterns to be employed as a biometric
identifier. We also investigated a computationally simpler method of matching such knuckle
patterns using local features and achieved ou tperforming results. The experimental results
presented in this paper also suggest that the combination of simult aneously acquired/ available
major, first minor and second minor knuckle patterns can achieve superior performance which is
not possible from any of the three finger knuckle patterns alone. The work detailed in this paper
also investigated the potential for visible illumination palm dorsal images as a biometric
identifier. Automatically segmented images from 501 subjects, with significant majority of them
acquired under outdoor illumination, were used to ascertain matching capability from such
potential identifier and encouraging results were obtained. Our results also demonstrated that the
combination of finger knuckle patterns and simultaneously extracted palm dorsal regions can be
used to further improve knuckle matching performance. The results presented from these set of
experiments should be consid ered preliminary, indicating great potential for this region to serve
as biometric, and require further work to achieve more accurate performance.
PROPOSED SYSTEM:
The most straightforward classifier in the classification or machine learning techniques is the
SVM classifiers ANN. Nearest Neighbour Classifier classification is achieved by identifying
the nearest neighbours to a query example and using those neighbours to determine the class of
the query. This approach to classification is of particular importance today because issues of
poor run-time performance are not such a problem these days with the computational power that
is available. This paper presents an overview of techniques for Nearest Neighbour classification
focusing on; mechanisms for assessing similarity (distance), computational issues in identifying
nearest neighbours and mechanisms for reducing the dimension of the data. The edge preserving
KNN based algorithm is proposed to improve overall performance and accuracy in knuckle
pattern matching.
 SOFTWARE USED

Introduction

MATLAB is a high-performance language for technical computing. It integrates

computation, visualization, and programming in an easy-to-use environment where problems and
solutions are expressed in familiar mathematical notation. Typical uses include the following:

Math and computation

Algorithm development

Data acquisition

Modeling, simulation, and prototyping

Data analysis, exploration, and visualization

Scientific and engineering graphics

Application development, including graphical user interface building

MATLAB is an interactive system whose basic data element is an array that does not
require dimensioning. This allows formulating solutions to many technical computing problems,
especially those involving matrix representations, in a fraction of the time it would take to write
a program in a scalar non-interactive language such as C or Fortran.

The name MATLAB stands for matrix laboratory. MATLAB was written originally to
provide easy access to matrix software developed by the LINPACK and EISPACK projects.
Today, MATLAB engines incorporate the LAPACK and BLAS libraries, constituting the state of
the art in software for matrix computation. In university environments, MATLAB is the standard
computational tool for introductory and advanced courses in mathematics, engineering, and
science. In industry, MATLAB is the computational tool of choice for research, development,
and analysis. MATLAB is complemented by a family of application specific solutions called
toolboxes.
Numerical Computing MATLAB

MATLAB (matrix laboratory) is a numerical computing environment and fourth-generation

programming language. Developed by MathWorks, MATLAB allows matrix manipulations,
plotting of functions and data, implementation of algorithms, creation of user interfaces, and
interfacing with programs written in other languages, including C, C++, and Fortran.
Although MATLAB is intended primarily for numerical computing, an optional toolbox
uses the MuPAD symbolic engine, allowing access to symbolic computing capabilities. An
additional package, Simulink, adds graphical multi-domain simulation and Model-Based Design
for dynamic and embedded systems.
In 2004, MATLAB had around one million users across industry and academia.
MATLAB users come from various backgrounds of engineering, science, and economics.
MATLAB is widely used in academic and research institutions as well as industrial enterprises.

INTERFACING WITH OTHER LANGUAGES

MATLAB can call functions and subroutines written in the C programming language or
Fortran. A wrapper function is created allowing MATLAB data types to be passed and returned.
The dynamically loadable object files created by compiling such functions are termed "MEX-
files" (for MATLAB executable).
Libraries written in Java, ActiveX or .NET can be directly called from MATLAB and
many MATLAB libraries (for example XML or SQL support) are implemented as wrappers
around Java or ActiveX libraries. Calling MATLAB from Java is more complicated, but can be
done with MATLAB extension,[ which is sold separately by MathWorks, or using an
undocumented mechanism called JMI (Java-to-Matlab Interface), which should not be confused
with the unrelated Java Metadata Interface that is also called JMI.
As alternatives to the MuPAD based Symbolic Math Toolbox available from
MathWorks, MATLAB can be connected to Maple or Mathematica.
MATLAB has a direct node with modeFRONTIER, a multidisciplinary and multi-
objective optimization and design environment, written to allow coupling to almost any
computer aided engineering (CAE) tool. Once obtained a certain result using Matlab, data can be
transferred and stored in a modeFRONTIER workflow and viceversa.

References:
[1] Y. Tang and N. Srihari, Likelihood ratio estimation in forensic identification using
similarity and rarity, Pattern Recognition, vol. 47, no. 3, pp. 945-958, Mar. 2014.
[2] A. Kumar and Ch. Ravikanth, Personal authentication using finger knuckle surface, IEEE
Trans. Info. Forensics & Security, vol. 4, no. 1, pp. 98-110, Mar. 2009
[3] AADHAR - Communicating to a billion, Unique Identification Authority of India, An
Awareness and Communication Report, ACSAC, http:/ / uidai.gov.in/ UID_PDF/
Front_Page_Articles/ Events/ A ADHAAR_PDF.pdf
[4] A. Kumar and Y. Zhou, Human identification using finger images, IEEE Trans. Image
Processing, vol. 21, pp. 2228-2244, April 2012.
[5] S. Ribaric and I. Fratric, A Biometric identification system based on eigenpalm and
eigenfinger Features, IEEE Trans. Pattern Anal. Mach. Intell., vol. 27, no. 11, Nov. 2005.
[6] A. Kumar and Y. Zhou, Human identification using knuckle-codes, Proc. 3rd Intl. Conf.
Biometrics, Theory and Applications, Washington D. C., BTAS'09, pp. 147-152, Sep. 2009.
[7] K. Sricharan, A. Reddy and A. G. Ramakrishnan, Knuckle based hand correlation for user
verification, Proc. SPIE vol. 6202, Biometric Technology for Human Identification III, P. J.
Flynn, S. Pankanti (Eds.), 2006. doi: 10.1117/ 12.666438
[8] M. Chora and Rafa Kozik, Contactless palmprint and knuckle biometrics for mobile
devices, Pattern Analysis & Applications, vol. 1, no. 15, 2012.
[9] L.-q. Zhu and S.-y. Zhang, Multimodal biometric identification system based on finger
geometry, knuckle print, and palm print, Pattern Recogn. Letters, vol. 31, pp. 1641-1649, Sep.
2010.
[10] L. Zhang, L. Zhang, D. Zhang and H. Zhu, Online fingerknuckle-print verification for
personal authentication, Pattern Recognition, vol. 43, no. 7, pp. 2560-2571, Jul. 2010
[11] T. Ojala, M. Pietikinen, and T. Menp, Multiresolution gray-scale and rotation
invariant texture classification with local binary patterns, IEEE Trans. Pattern Anal. Mach.
Intell., vol. 24, no. 7, pp. 971-987, 2001.
[12] H. Jin, Q. Liu, H. Lu, X. Tong, Face detection using improved LBP under Bayesian
framework, Proc. ICIG, pp. 306-309, Dec. 2004.
[13] Q. Zheng, A. Kumar, G. Pang, Suspecting less Suspecting less and achieving more: New
insights on palmprint identifica tion for faster and more accurate palmprint matching, IEEE
Trans. Information Forensics & Security, vol. 11, pp. 633-641, Mar. 2016.
[14] D. G. Joshi, Y. V. Rao, S. Kar, V. Kumar, and R. Kumar, Computer vision based approach
to personal identification using finger crease patterns, Pattern Recognition, vol. 31, pp. 15-22,
Jan. 1998. [15] D. L. Woodard and P. J. Flynn, Finger surface as a biometric identifier,
Computer Vision and Image Understanding, vol. 100, no. 3, pp. 357-384, Dec. 2005.
[16] A. Kumar, Can we use minor finger knuckle images to identify humans?, Proc. BTAS
2012, pp. 787-87g, Sep. 2012.
[17] K. Ito, T. Aoki, H. Nakajima, K. Kobayashi, T. Higuchi, A palmprint recognition
algorithm using phase-only correlation, IEICE Trans. Fundamentals of Electronics,
Communications and Computer Sciences, vol. E91-A, pp. 1023-1030, April 2008.
[18] Z. Sun, T. Tan, Y. Wang, S. Z. Li, Ordinal palmprint representation for personal
identification, Proc. CVPR 2005, pp. 279-284, 2005.
[19] W. Jia, D.-S. Huang, D. Zhang, Palmprint verification based on robust line orientation
code, Pattern Recognition, vol. 41, pp. 1504-1513, May 2008.
[20] P. Sinha. Qualitative representations for recognition, Biologically Motivated Computer
Vision, pp 249--262. Springer, 2002.