Matching Sketches with Digital Face Images using MCWLD and Image Moment Invariant

IOSR Journal of Computer Engineering (IOSR-JCE)
e-ISSN: 2278-0661,p-ISSN: 2278-8727, Volume 17, Issue 6, Ver. I (Nov – Dec. 2015), PP 131-137
www.iosrjournals.org
DOI: 10.9790/0661-1761131137 www.iosrjournals.org 131 | Page
Matching Sketches with Digital Face Images using MCWLD and
Image Moment Invariant
Reshma C Mohan1
, M. Jayamohan2
, Arya Raj S3
1
(PG Scholar, Sree Buddha College of Engineering for Women, Elavumthitta, Kerala, India)
2
(Head of Section, College of Applied Science, Adoor, Kerala, India)
3
(Assistant Professor, Sree Buddha College of Engineering for Women, Elavumthitta, Kerala, India)
Abstract : Face recognition is an important problem in many application domains. Matching sketches with
digital face image is important in solving crimes and capturing criminals. It is a computer application for
automatically identifying a person from a still image. Law enforcement agencies are progressively using
composite sketches and forensic sketches for catching the criminals. This paper presents two algorithms that
efficiently retrieve the matched results. First method uses multiscale circular Weber’s local descriptor to encode
more discriminative local micro patterns from local regions. Second method uses image moments, it extracts
discriminative shape, orientation, and texture features from local regions of a face. The discriminating
information from both sketch and digital image is compared using appropriate distance measure. The
contributions of this research paper are: i) Comparison of multiscale circular Weber’s local descriptor with
image moment for matching sketch to digital image, ii) Analysis of these algorithms on viewed face sketch,
forensic face sketch and composite face sketch databases.
Keywords - composite sketches, image moment, support vector machines, weber’s local descriptor
I. Introduction
There are three types of sketches: forensic sketches, viewed sketches and composite sketches. Forensic
sketches are the sketches drawn by specially trained artists based on the description of criminal by an eye
witness. Viewed sketches are the sketches drawn by an artist, directly looking at the subject or the photograph of
the criminal. Composite sketches [7] are drawn using software tools that facilitate an eyewitness to select
different facial components from a wide range of pre-defined templates. The eyewitness selects the most
resembling facial template for each feature based on his/her recollection from the crime scene. These tools allow
processing each feature individually and then combine all the features to generate a composite sketch.
Compared to hand drawn sketches the composite sketches require lesser effort both in terms of cost as well as
time. Fig 1. shows the example of viewed, forensic and composite sketches and its corresponding photograph.
Matching sketches with digital face images is a very important law enforcement application. Developments in
biometric technology have provided additional tools to criminal investigators that help to determine the identity
of criminals [1].
(a) (b) (c)
Fig 1. (a)Viewed sketch and it's corresponding photograph, (b) Forensic sketch and it's corresponding photograph,
(c) Composite sketch and it's corresponding photograph.
If a fingerprint is found at the scene of crime or a surveillance camera captures an image of the face of
a suspect, these are used in determining the suspect using different biometric identification techniques. In many
cases such type of above information is not present. The technology to efficiently capture the biometric data like
finger prints within a short period of time is impractical. So in many cases an eyewitness is available in crime
who had seen the criminal. The forensic artist draw sketch that portrays the facial appearance of the criminal.
Sketches drawn by using such process is called as forensic sketches. When sketch is ready, it is sent to the law
enforcement officers and media outlets with the hope of catching the suspect. Generally, these sketches are
manually matched with the database consisting of digital face images of known individuals. If the criminal has
been imprisoned at-least once, a mug shot photo (photo taken, while the person is being sent to jail) is available.

Matching Sketches with Digital Face Images using MCWLD and Image Moment Invariant
Using an efficient sketch matching system, the police can narrow down the potential suspects which will reduce
the future crimes by the same criminal. Fig 2. shows the representation of sketch matching system.
The Weber’s local descriptor [2] is based on Weber’s law and it is similar to Scale Invariant Feature
Transform (SIFT) and Local Binary Pattern (LBP). It is similar to LBP in analyzing small neighborhood regions
and to SIFT in computing histogram using gradient magnitude and orientation. WLD has some unique features
that compute the salient micro patterns in a relatively small neighborhood region.
Fig 2. Representation of Sketch Matching System
II. Related Works
Jiang, et al. [3] proposed Scale Invariant Feature Transform (SIFT) feature based face recognition.
SIFT features are features extracted from images in reliable matching between different views of the same
object. That extracted features are invariant to scale and orientation. After computing the SIFT features the
common representation for sketch image and for photo is found out. To compare the similarity between photo
and sketch, directly compare the distance between the two common representation vectors. The SIFT features
described in this implementation are computed at the edges and they are invariant to image rotation, scaling and
addition of noise.
Ambhore et al. [4] proposed face sketch to photo matching using Local Feature-Based Discriminant
Analysis (LFDA). In LFDA framework [5], scale invariant feature transform (SIFT) and multiscale local binary
pattern (MLBP) feature descriptors are used. Minimum distance sketch matching can be performed using this
feature-based representation by computing the normed vector distance. The local binary patterns (LBP) operator
takes a local neighborhood around each central pixel, thresholds the neighborhood pixels and uses resulting
binary-valued image patch as a local image descriptor.
Balaji et al. [6] Extended Uniform Circular Local Binary Pattern (EUCLBP) Matching Algorithm. It
extracts discriminating information present in local facial regions at different levels of granularity. Digital face
images and sketches are decomposed into multi resolution pyramid which forms the discriminating facial
patterns. EUCLBP use these patterns to form a unique signature of the face. For matching, a memetic
optimization based approach is used to find the optimum weights corresponding to each facial region.
Chen et al. [2] proposed Weber’s Local Descriptor (WLD) inspired by Weber’s law. It states that the
change of a stimulus that will be just noticeable is a constant ratio of the original stimulus. Weber's Local
Descriptor is a dense descriptor computed for every pixel and depends on both the magnitude of the center pixel
intensity and the local intensity variation. 2D WLD histogram is used for texture classification. There are two
WLD components, differential excitation and orientation which is used to construct a WLD histogram.
Bhatt et al. [1] proposed Memetically Optimized Multi Scale WLD (MCWLD). Multi-scale Circular
WLD is one of the most advanced types of image descriptors. WLD is optimized for matching sketches with
digital face images by computing the multi-scale descriptor in a circular manner. Memetic optimization
algorithm is used to assign optimal weights to every local facial region to boost the identification performance.
This algorithm extracts the discriminating information from local regions of both digital face images and
sketches.
Hu [8] proposed Regular Image Moment Invariant based matching sketch with digital image. Regular
image moment invariants are invariant to rotation, scaling and translation. It provides information such as
localized weighted average of the pixel intensities, centroid, and orientation details of an image. Image moments
can efficiently encode selective information around the important features in local regions of sketches and
digital face images. There are 7 image moments. Generally, each moment indicates a specific function. The first
moment is the mean, the second moment is the variance and so on.

III. Proposed Work
3.1 Support Vector Machine for Preprocessing
There are various methods for image classification. SVM is one of the best known methods in pattern
classification and image classification. Given a set of training examples, each marked for belonging to one of
two categories, an SVM training algorithm assigns new examples into one category or the other. Most of
classifiers, such as neural network, decision tree, minimum distance, maximum likelihood and support vector
machine require a training sample. Based on the training set the images are classified into a set of matched and
unmatched images. The sketch-digital image pair that is classified based on this is considered for the further
experimentation.
3.2 Matching sketches with Digital Face Images using MultiScale Circular Weber’s Local Descriptor
Chen et.al [2] proposed Weber’s Local Descriptor (WLD) which is based on Weber’s law. Weber’s law is
expressed as:
(1)
Where ∆I represents the increment threshold, I represents the initial stimulus intensity and k shows that
the proportion on the left side of the equation remains constant despite of variations in the I term.
WLD computes the salient micro patterns in a relatively small neighborhood region with linear
granularity [1]. WLD is used for matching sketches with digital face images by computing multiscale descriptor
in circular manner. The multiscale circular WLD histograms of both sketch and digital face image are matched
using weighted distance measure.
The given face image is converted into gray scale image and it is resized into an adequate dimension.
Multiscale Circular WLD (MCWLD) descriptor is computed for different values of P and R. P is the number of
neighboring pixels on a circle of radius R centered at current pixel. For every pixel, the multiscale analysis is
performed by varying the number of neighbors P and radius R (R=1, 2, 3 and P=8, 16, 24). The fig 3. represents
the steps involved in matching system. MCWLD has two components differential excitation and orientation.
The sketches and digital face images are represented using MCWLD as follows.
Differential Excitation: For every pixel, differential excitation [1] is computed as the arctangent
function of ratio of the intensity difference between the central pixel and its neighbors to the intensity of the
central pixel. The differential excitation of central pixel ξ ) is computed as:
(2)
where is the intensity value of the central pixel and P is the number of neighbors on a circle of radius R.
Orientation: Orientation component computed as:
(3)
The orientation value is quantized into T dominant orientations. T is experimentally set as eight. Before
the quantization, the orientation value θ is mapped into the interval [0,2Π] to make the θ value in that range.
Therefore, the quantization function is as follows:
(4)
Here T=8, therefore (t=0, 1…..T-1).

Fig 3. Steps involved in matching sketches.
MCWLD Histogram: We first compute the CWLD features for each pixel (i.e., CWLD , where j=0,1..N-
1, t=0,1….T-1 and N=is the dimension of the image). The differential excitation are regrouped into T
orientation sub-histograms H (t) (t=0,1…T-1). Each sub-histogram H (t) corresponds to a dominant orientation.
Within each dominant orientation, the range of differential excitation (i.e., sub-histogram H (t)) is divided into
M intervals (m=0,1….M-1 and M=6). The differential excitation value for each pixel is then differentiated
on the basis of the eight different quantum levels. These differential excitation values corresponding to first
quantum level are assigned to h1, differential excitation values corresponding to the second quantum level are
assigned to h2 and so on. Fig 4. shows the steps involved in computing the circular WLD histogram.
For each differential excitation intervals (n=0,1….N-1), the lower bound and upper bound is computed and
this range is divided into M=6 segments. These sub-histogram segments across all dominant orientations
are reorganized as M one dimensional histograms. Then the M sub-histograms are concatenated into a single
histogram representing the circular WLD histogram. The multiscale circular WLD is computed with different
values of P and R. Multiscale analysis is performed in three different scales i.e., (R=1, P=8), (R=2, P=16) and
(R=3, P=24). These histograms at different scales are concatenated to form the facial representation.
Figure 4. Illustrating the steps involved in computing the circular WLD histogram [2].

Weight for MCWLD Histogram: The MCWLD histogram represents the frequency information. Some facial
regions are more discriminating than others and it will contribute more for recognition. The regions with high
variance are more discriminating than flat regions. Thus, this M segments play different roles for
recognition. For better performance, different weights need to be assigned to diff histogram or frequency
segments. Here we are taking weight as one for accurate result. Weighted measure is used for matching two
MCWLD histograms. It is represented as [1]:
(5)
where p and q are the two MCWLD histograms to be matched, i and j correspond to the bin of the
histogram segment and is the weight of the histogram segment. Here is taken as one based on
the experimentation.
3.3 Matching Sketches with Digital Face Images using Image Moment Invariant
Hu [8] derived a set of 7 moment invariants, which are the properties of connected regions in binary
images that are invariant to translation, rotation and scale. Image moment is a particular weighted average of the
image pixel intensities. Function of such moments, usually chosen to have some attractive property or
interpretation. It can efficiently encode the selective information around the prominent features in local regions
of sketches and digital face images. Out of the 7 image moments proposed by Hu [8] second image moment
is determined to be the most selective feature for representing the shape and orientation of local facial regions in
sketches. Second image moment is computed as:
(6)
where represents normalized central moment. It is represented as:
(7)
where is the central moment. The central image moments are regular two dimensional image moments
which are shifted from the origin to image centroid . Therefore, , central moment is calculated as:
(8)
where , . The two-dimensional regular moments of order is computed as:
(9)
where is the intensity value at and p, q = 0,1,2….n.
The second order image moment is for each pixel in the image is calculated using Equations (6-9). Then the
feature image is tessellated into 3x3 non-overlapping patches. Histogram is constructed for each local patch. The
histogram that represents facial representation is computed where every uniform pattern has a separate bin and
each non-uniform pattern is assigned to a single bin. The complete image signature is obtained by concatenating
all the histograms of ach local patches. To compare a gallery-probe pair, the distance measure is used.
(10)
where x and y are the two feature histograms to be matched, i is the local region and j is the histogram bin.
IV. Experimental Analysis
The performance of MCWLD compared with image moment invariant algorithm for matching sketches
with digital face images. For analysis, forensic sketches, composite sketches and viewed sketches are used.

4.1 Accuracy
The performance of proposed and existing algorithm designed for matching sketches with digital face
images is compared. The preprocessing classification reduces the possibility of false positives and the number of
comparisons. By preprocessing only top matched digital face images are selected for the next step. i.e., in the
case of most sketches the exact match will be in the first 50%.
The performance of proposed and existing algorithm designed for matching sketches with digital face
images is compared. The preprocessing classification reduces the possibility of false positives and the number of
comparisons. By preprocessing only top matched digital face images are selected for the next step. i.e., in the
case of most sketches the exact match will be in the first 50%.
The databases used in the work are IIIT-Delhi database [9], AR database[10] and PRIP composite
sketch database[11]. Composite sketches for each subject are synthesized using two commercial facial
composite systems: (i) FACES [12], and (ii) IdentiKit [13]. The AR database and IIIT-D viewed sketch database
have variations introduced by disparate drawing styles of artists. As discussed by Zhang et.al [14] drawing
styles of different artists play an substantial role in how closely a sketch resembles the actual digital photo thus
influencing the performance of automatic algorithms. In viewed sketches, the sketch is closely similar to digital
face images, then the identification accuracy of matched results will be high.
Previous research [15] in matching forensic sketches suggests that existing sketch recognition
algorithms trained on viewed sketches are not sufficient for matching forensic sketches. Further it degrade the
performance of sketch to digital image matching algorithms. Since the forensic sketches are based on the
recollection of the eyewitness, they are often incomplete, inaccurate, do not closely resemble the actual digital
face images. So the identification accuracy of matched results will be low. But the proposed work will increase
the identification accuracy.
Again the performance of the proposed is evaluated on the PRIP database. Composite sketches may not
include minute feature details as compared to what an artist can represent in hand drawn sketches, therefore
composite sketches often look synthetic. Matching composite sketches with digital face images shows an even
more challenging problem for automatic face recognition algorithms than matching hand-drawn sketches with
digital face images. But the proposed work increases the identification accuracy of matched results.
Table 1 shows the identification accuracy of both MCWLD and Image moment invariants. Compared
to MCWLD the proposed scheme improves the identification accuracy in the case of forensic and composite
sketches.
Table 1. Identification Accuracy (%) Of Sketch To Digital Face Image Matching Algorithms MCWLD
And Image Moment Invariant For Matching Viewed Sketches, Forensic Sketches And Composite
Sketches
Algorithm Viewed
Sketches
(Accuracy %)
Forensic
Sketches
(Accuracy %)
Composite
Sketches
(Accuracy %)
MCWLD 97.2 64.37 84.61
Image Moment
Invariant 97.2 81.9 92.3
4.2 Time Estimation
Matlab Version: MATLAB 7.12.0
Processor: Intel Core i3
RAM: 6.00GB
System Type: 64-bit Operating System
The execution time of both the works is analyzed. The Table 2 shows the execution time of both
MCWLD and Image moment invariant in the above system. From the analysis it is clear that MCWLD take
double execution time in retrieving results as compared to the Image Moments.
Table 2. Total Execution Time Of Sketch To Digital Face Image Matching
Algorithms MCWLD And Image Moment Invariant
Number of Images MCWLD Image Moment Invariant
10 0.678 0.319s
20 1.338 0.624s
100 10.704 5.328m

V. Conclusion
This research focuses on matching sketches with digital face images. The classification is used in this
work because of two reasons: (1) to avoid the number of false positives and (2) to increase the computational
efficiency i.e., it will reduce the number of comparisons. In the proposed algorithm, discriminative information
is extracted from local facial regions using second order image moments. It provides information such as
localized weighted average of the pixel intensities, centroid and orientation details of an image. Second moment
is the most selective feature for representing shape and orientation of local facial regions. In the analysis
including the comparison between MCWLD and IMI, is performed on using the viewed, forensic and composite
sketch databases. The results show that the proposed IMI algorithm is significantly better than existing
MCWLD. As a future work, we plan to include weight optimization with IMI to assign optimal weight to every
local facial region. The memetic optimization can be used in weight optimization and dimensionality reduction.
References
[1] Himanshu S. Bhatt, Samarth Bharadwaj Richa Singh, and Mayank Vatsa, ‖Memetically Optimized MCWLD for Matching Sketches
With Digital Face Images,‖ in Proc. IEEE Transactions on information forensics and security, vol. 7, no. 5, october 2012, pp.1522-
1535.
[2] J. Chen, S. Shan, C. He, G. Zhao, M. Pietikainen, X. Chen, and W. Gao, ―WLD—A robust local image descriptor,‖ IEEE Trans.
Pattern Anal. Mach. Intell., vol. 32, no. 9, pp. 1705–1720, Sep. 2010.
[3] C. Geng and X. Jiang, ‖Face recognition using SIFT features,‖ in Proc. Int. Conf. Image Processing, 2009, pp. 3313-3316.
[4] Pushpa Gopal Ambhore and Lokesh Bijole, ‖Face Sketch to Photo Matching Using LFDA,‖International Journal of Science and
Research (IJSR), 2014, pp.11-14.
[5] Surya Raj, and Navya Jose, ‖Digital Face Recognition Techniques for Matching with Sketches,‖ in Proc. IOSR Journal of
Electronics and Communication Engineering, vol. 9, issue 2, Ver. IV (Mar - Apr. 2014), pp.20-36.
[6] Gnanasoundari.V and Balaji.S, ‖EUCLBP Matching Algorithm for Forensic Applications,‖ in Proc. International Journal of
Emerging Technology and Advanced Engineering, Volume 4, Issue 2, February 2014, pp.121-126.
[7] Scott Klum, Hu Han, Anil K. Jain and Brendan Klare, "Sketch Based Face Recognition: Forensic vs. Composite Sketches",
Michigan State University, East Lansing, MI, U.S.A.
[8] M.-K. Hu, "Visual pattern recognition by moment invariants". IRE Transactions on Information Theory, 8(2):179–187, 1962.
[9] H. Bhatt, S. Bharadwaj, R. Singh, and M. Vatsa, On matching sketches with digital face images,‖ in Proc. Int. Conf. Biometrics:
Theory Applicatons and Systems, Washington, DC, 2010.
[10] A. Martinez and R. Benevento, The AR Face Database CVC Tech. Rep. 24, 1998.
[11] H. Han, B.F. Klare, K. Bonnen, and A.K. Jain, ―Matching composite sketches to face photos:a componentbased approach,‖ IEEE
Transactions on Information Forensics and Security, vol. 8, no. 1, pp. 191–204, 2013.
[12] FACES: The ultimate composite software. Available at-http:// www.facesid.com.
[13] Identi-Kit, Identi-Kit Solutions, http://www.identikit.net/, 2011
[14] J. S. Y. Zhang, C. McCullough and C. Ross, " Hand-drawn face sketch recognition by humans and a pca-based algorithm for
forensic applications," IEEE Trans. Syst., Man, Cybern. A, Syst., Humans, vol.40, no.3, pp. 475-485, May 2010.
[15] L. Z. B. Klare and A. Jain, "Matching Forensic Sketches to Mug shot Photos,_ IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no.
3, pp. 639-646, Mar. 2011.

Matching Sketches with Digital Face Images using MCWLD and Image Moment Invariant

More Related Content

What's hot (16)

Viewers also liked (10)

Similar to Matching Sketches with Digital Face Images using MCWLD and Image Moment Invariant (20)

More from iosrjce (20)

Recently uploaded (20)

Matching Sketches with Digital Face Images using MCWLD and Image Moment Invariant