2538 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO.
10, OCTOBER 2018
Online Writer Identification With Sparse
Coding-Based Descriptors
Vivek Venugopal and Suresh Sundaram
Abstract— This paper proposes a system to identify the author-
ship of online handwritten documents. We represent the trace of
handwriting of a writer with descriptors that are derived from a
set of dictionary atoms obtained in a sparse coding framework.
The descriptors for each dictionary atom encode the error while
using it alone for reconstruction. The use of sparse representation
offers flexibility in describing each of the segmented handwritten
sub-strokes of a writer as a combination of more than one
atom or prototype. The descriptor is constructed by considering
the attributes obtained from sets of histograms extracted at a
sub-stroke level. In addition, an entropy-based analysis for the
bin size to be used for obtaining the feature sets is proposed. The
writer descriptor is evaluated on the paragraph and individual
text lines of two publicly available English databases (the IAM
and IBM-UB1) and a Chinese database—CASIA. We empirically
show that the results obtained are promising when compared with
previous works.
Index Terms— Online writer identification, substrokes, his-
togram based features, sparse coefficients, writer descriptor,
IAM online handwriting database, IBM-UB 1 database, CASIA
database.
I. I NTRODUCTION
R ESEARCH works on writer identification focus on the
task of assigning authorship of a piece of handwritten
document by contrasting it against templates of handwritten
samples saved in a database [1]. In literature, these come under
the purview of behavioral biometrics. Past explorations in this
area have led to the development of two types of systems,
namely online and offline.
In an online writer identification system, the handwritten
input is captured by an electronic stylus that makes contact
with the screen of a tablet. The tip of the stylus records the
dynamic information of the trace such as (x, y) coordinates
and time stamp. In general, the input to such a system is a
document comprising a set of strokes, that in turn contains a
sequence of points. Offline writer identification, on the other
hand, regard the handwritten data as an image and consider
techniques from image processing for subsequent analysis.
In such systems, there are two categorizations, namely textural
based and allograph based methods. In the former group,
Manuscript received August 21, 2017; revised December 29, 2017,
February 25, 2018, and March 12, 2018; accepted March 16, 2018. Date
of publication April 4, 2018; date of current version May 9, 2018. The
associate editor coordinating the review of this manuscript and approving it for
publication was Dr. Julian Fierrez. (Corresponding author: Suresh Sundaram.)
The authors are with the Department of Electronics and Elec-
trical Engineering, IIT Guwahati, Guwahati 781039, India (e-mail:
v.venugopal@iitg.ernet.in; sureshsundaram@iitg.ernet.in).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIFS.2018.2823276
1556-6013 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
global description from the handwritten image data are uti-
lized for writer identification such as computation of ink
width [2] and attributes like Local Binary Pattern (LBP) [3],
Co-occurrence features [4] and Edge structure code [5] to
name a few. Conversely, the allograph based techniques con-
sider the feature description computed from the letter parts
(allographs). While establishing the authorship of a test docu-
ment, these entail the use of a dictionary / vocabulary that are
pre-learnt during the time of training [6]–[8]. In the context of
the present paper on online writer identification, the allographs
relate to the sub-strokes that are segmented from the trace of
the handwritten input.1
The research ideology of this paper falls in line with the
allograph-based method. Without loss of generality, works
from such approaches in literature suggest the generation of
the vocabulary via a grouping of the sub-strokes by a hard
clustering algorithm like k-means. Such a scheme allows each
feature vector of a segmented sub-stroke (from the input data
of a writer) to be associated or represented by exactly one
code-vector in the vocabulary. However, many a times, it may
not be possible to represent sub-strokes with only one code-
vector. To resolve this issue, we adopt a sparse framework
for constructing the dictionary. This provides flexibility to
describe each of the segmented sub-strokes of a writer as
a linear combination of more than one atom / prototype.
In recent times, sparse representation has been applied very
successfully in various applications of pattern analysis such as
face recognition [9], [10], signature verification [11], to name
a few.
A. Focus of the Present Work
This paper focuses on proposing a methodology that
encodes the sub-strokes of the online handwritten trace
with descriptors derived from the set of dictionary atoms.
We attempt to capture the similarity of the attributes2 of
each individual sub-stroke to the corresponding value in the
subset of atoms that contribute with a non-zero sparse coding
coefficient. Keeping this in perspective, we derive descriptors
from each of the atoms that incorporate the similarity scores of
the attributes in a feature vector. The scores, as we shall see
in Section VI, are in a way indicative of the reconstruction
1 The textural based methods rely on the bitmap image and hence are not
applicable for the categorization of online writer identification systems. The
previous works from literature relate primarily to the use of allograph based
methods and these are discussed in Section II.
2 We refer to the feature values in a feature vector as ‘attributes’. The
attributes computed for each sub-stroke of the online trace are stacked to
form a feature vector.
VENUGOPAL AND SUNDARAM: ONLINE WRITER IDENTIFICATION WITH SPARSE CODING-BASED DESCRIPTORS
Fig. 1. Block diagram of the proposed text independent online writer identification methodology. The test data refer to paragraphs or text lines written by
the enrolled writers.
error obtained while employing a particular atom alone for
reconstruction. The idea of scoring individual feature attributes
separately with regards to a dictionary atom is achieved by
utilizing the corresponding non-zero sparse coefficient.
Prior attempts on sparse coding primarily rely on the
accumulation of the coefficients over the set of dictionary
atoms for generating the descriptor [6], [12]. Unlike these
works, the present proposal takes a step forward in exploring
additional information from the sparse coefficients for writer
identification. Their utilization in separately scoring each of
the individual attributes in a feature vector helps in capturing
the description of the writer at a finer level, compared to the
works [6], [12]. This in turn leads to improvements in perfor-
mance of the writer identification system.
In addition to the above contribution, we consider in
Section IV, several sets of histogram based attributes / features
at the sub-stroke level for constructing the dictionary atoms
and subsequently the descriptors.3 We provide a more thor-
ough and comprehensive analysis of the same by addressing
the issue of the selection of the bin size to be used and
their relation to the writer identification rate. More specifically,
we propose in Section IV-A, an entropy based strategy for
determining the appropriate bin size to be selected for the
generation of these histograms. Considering that the descrip-
tors of the sparse coding framework are constructed from the
resulting histogram feature sets, the selection of the bin size
using our method ensures a good discrimination between the
writers. This also gets reflected in the experimental results,
where we show a dependence between the bin size chosen
by the entropy based analysis and the corresponding writer
identification performance.
The efficacy of our proposal is demonstrated with extensive
experimentations on three publicly available online handwrit-
ten databases namely IAM, IBM-UB1 and CASIA. Apart from
the one-versus-all SVM, we also consider the Exemplar-SVM
framework discussed in [8] for writer identification.
We summarize the discussion of this sub-section by reiter-
ating the contributions of the paper.
3 These features were introduced for the first time by us in the paper [12]
published in ICFHR-2016.
2539
1) We provide a more thorough analysis of the histogram
based features introduced in [12]. A strategy based on
entropy is proposed in subsection IV-A for determining
the appropriate bin size to be selected for the histograms.
2) We present a proposal for describing the handwritten
content by a set of descriptors derived from the set of
dictionary atoms in subsection VI.
3) Lastly, extensive evaluation of the proposal is presented
on two English databases (IAM and IBM-UB1) and a
Chinese database (CASIA) with promising results over
several state-of-art methods.
To the best of our knowledge, the entropy based analysis for
bin size selection of the histogram based features as well as
the derivation of the descriptor from the dictionary atoms
has not been attempted in the literature on online writer
identification.
B. A Block Schematic of Our Proposal
Figure 1 presents the block diagram encapsulating the
details of our proposal. The spatio-temporal sequence con-
sisting of the spatial coordinates and the time duration from a
document is initially passed through the pre-processing block.
The output of this module is then used to segment the strokes
into smaller fragments called sub-strokes, following which
histogram based feature sets parameterized by a bin size are
extracted for each of them. The atoms of a pre-learnt dictionary
are then used in conjunction with the attributes of sub-stroke
feature vectors to generate a descriptor, that encapsulates
the writer specific characteristics of the document. Finally,
a Support Vector Machine (SVM) classifier is utilized to
establish the authorship of the handwritten content.
The organization of this paper is as follows: We begin by
reviewing the allograph based methods from the literature
of online writer identification in Section II. Following this,
we outline in the subsequent four sections the steps of our
proposed writer identification. The pre-processing methods
together with the sub-stroke generation process are presented
in Section III. The histogram based feature sets along with an
entropy based analysis (for the appropriate bin size selection)
is described in Section IV. An overview of the sparse coding
2540 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO. 10, OCTOBER 2018
framework for online writer identification with appropriately
defined notations is then provided in Section V. In Section VI,
the sub-stroke feature vectors from a document together with
their corresponding sparse-coded coefficients are utilized in
deriving the writer descriptor.
With regards to the experiments, the set of databases are
discussed in Section VII with their respective training and
testing protocols. The efficacy of our proposal is demon-
strated in Section VIII, followed by a discussion to prior
works in Section IX. Lastly, in Section X, we conclude the
paper.
II. OVERVIEW OF A LLOGRAPH BASED A PPROACHES IN
O NLINE W RITER I DENTIFICATION
Over the last decade, there has been considerable research
in the area of online writer identification under the category
of allograph-based approaches. One of the prominent works
is that of the Gaussian Mixture Model-Universal Background
Model (GMM-UBM) system proposed in [13] and [14]. The
idea is to first construct the UBM from the training data
corresponding to a set of enrolled writers with the Expectation
Maximization (EM) algorithm. Thereafter, the adaptation step
is considered wherein individual GMMs are obtained for
each writer by using only his / her training data. During the
testing phase, the document is assigned to the writer, whose
corresponding GMM provides the highest log-likelihood score.
The development of online writer identification systems
have also been motivated from the domain of information
retrieval. The works in [15]–[17] consider scoring the dis-
tribution of the allographs in the handwritten data with a
measure motivated from the term frequency-inverse document
frequency (tf-idf) framework. In addition, the notion of Latent
Dirichlet Allocation from the area of topic models has been
explored in [18]–[20]. The underlying assumption made in
these works is that a document can be modelled as a distrib-
ution of finite writing styles shared among writers, which in
turn can be viewed as a random mixture of text independent
features.
Apart from the preceding ideas, a study towards the utility
of subtractive clustering to generate the prototype patterns for
online writer identification was conducted in [21]. In partic-
ular, two different strategies have been put forth for deciding
the identity of the writer. The first one makes use of a scoring
inspired from the tf-idf approach while the second utilizes
a voting scheme based on the nearest prototype. Following
the success of sparse representation approach in offline writer
identification [6], a study was conducted to demonstrate its
effectiveness for online handwritten documents in [12]. In this
work, the identification was achieved by employing the sparse
coefficients in a tf-idf framework. An investigation on the
viability of deriving writer description from a pre-learnt code-
book generated with k-means algorithm was presented in [22].
In particular, a comparative study was conducted with the
Vector of Locally Aggregated Descriptors [23] for identifying
the authorship of online handwritten documents.
Methods that make use of shape primitives have also been
attempted in the literature. Namboodiri and Gupta [24] encode
the size, shape and style of a specified set of shape primitives4
to deduce the identity of the writer. The work in [25] charac-
terises each stroke in a document by utilizing the histogram of
dynamic information such as pressure, azimuth, altitude and
velocity. The idea of shape based and temporal sequence codes
has also been attempted in the literature [26]. At this juncture,
we also make a mention of the hierarchical approach to writer
identification adopted in [27]. In the first stage, the statistics
of the orientation and the pressure information in a shape
primitive (captured as a probability distribution) are employed
to reduce the list of probable writers. In the second stage,
the curvature of the shape primitive of the selected writers is
quantized to 18 bins. Thereafter, the mean and variance of four
dynamic features (pressure, azimuth, altitude and velocity) in
each bin is employed for identification.
In a recent study, Chaabouni et al. [28] attempted to
model the sub-strokes using multi-fractals for Arabic writer
identification. The sub-strokes were clustered into five groups
based on the frequently occurring patterns in Arabic. In the
work [29], an approach based on the combination of Dynamic
Time Warping (DTW) and Support Vector Machine (SVM)
has been proposed for ascertaining the authorship of texts in
Arabic. An exploration towards the efficacy of Beta-Elliptical
models in representing the spatial and velocity profile of a sub-
stroke was suggested in [30]–[32] for writer identification.
We conclude with a mention of deep learning based
approaches, that have recently captured the interest of the
online writer identification research community. A Convolu-
tional Neural Network based approach was presented in [33]
where the authors employ a drop segment method (influenced
from the drop-out training strategy of [34]) to improve the
generalization capability. Lastly, a Recurrent Neural Network
model with bi-directional Long Short Term Memory based
approach has also been proposed in the work [35].
III. P REPROCESSING AND S UB -S TROKE G ENERATION
The following two preprocessing steps are employed in this
study.
• Isolated point removal: In order to eliminate such points,
we consider a neighbourhood Np around a sample point
(say p) in the same stroke. Thereafter, we compute the
Euclidean distances between p and all the other points
in Np . If all the distances are greater than a thresh-
old t, then it is regarded as an isolated point and sub-
sequently eliminated from the online handwritten trace.
The threshold is computed empirically for each document
by considering the mean μ and standard deviation σ of
the Euclidean distances calculated between consecutive
sample points of a stroke. For our work, the value of t is
set to (μ + 3σ ).
• Stroke merge: Due to issues in the capture of the online
data, at times, we encountered that a handwritten trace,
(which can be represented as a single stroke) is broken
down to more than one stroke.5 In such cases, we consider
4 In this paper, the terms shape primitives, sub-strokes and allographs have
been used interchangeably.
5 This issue was observed with regards to the IAM Online handwriting
database, that will be used for the experimentation in this work.
VENUGOPAL AND SUNDARAM: ONLINE WRITER IDENTIFICATION WITH SPARSE CODING-BASED DESCRIPTORS
Fig. 2. Illustration of the segmentation approach - (a) Input online handwritten trace. (b) Substrokes obtained by locating the local maxima in a stroke.
(c) Substrokes generated with considering fixed length sample points. Note that we use different colours to distinguish between subsequent sub-strokes.
the time difference between the last point of a stroke and
the first point of the succeeding stroke. If the obtained
value lies within a threshold, the corresponding strokes
are merged into a single stroke.
Once the preceding pre-processing steps are performed, each
stroke is divided into fragments called sub-strokes, that are
in a way indicative of the basic structural units present in
the document. The segmentation procedure utilized in this
work considers the generation of sub-strokes obtained via two
strategies as outlined below: In the first method, we rely on
locating the local maxima in a stroke. The strokes are then split
at these points and individual sub-strokes obtained thereof.
In the second strategy, we consider sub-strokes of a fixed
length by splitting the stroke into smaller segments containing
a specific number of points. The successive segments are
allowed to have an overlap that is governed by a stride
factor. For our work, we have considered a sub-stroke to
contain 30 points with a stride of five points. The generation of
sub-strokes using the segmentation strategies may be regarded
as a data augmentation process, that facilitates the learning
of the sparse representation framework on their corresponding
feature vectors discussed in the next section.
As an illustration, we depict in Figure 2 (a) an online
trace of a word and the sub-strokes obtained using the two
strategies in sub-figures (b) and (c) respectively. For better
clarity, different colours are employed to distinguish between
subsequent sub-strokes.
IV. F EATURE E XTRACTION
Subsequent to generating the sub-strokes, we extract fea-
tures that encode their shape information. Such a represen-
tation provides a fixed feature vector dimension irrespective
of the length of the sub-stroke. Inspired by the success of
Histogram of Oriented Gradient (HOG) [36] in the area of
image processing, we propose sets of histogram based features
/ attributes for each sub-stroke as follows:
NS
Consider a sub-stroke S of N S points denoted by {pi }i=1
where pi = (x i , yi ) consists of the spatial x and y coordinates
of the i t h point. Let pc = (x c , yc ) represent the coordinates
of its centroid. We calculate θpi , the angle of the line segment
pi − pc with the horizontal and the Euclidean distance dpi
between pi and pc as follows:
 
−1 (yi − yc )
θpi = tan
(x i − x c )

dpi = (x i − x c )2 + (yi − yc )2 (1)
2541
Once the (θpi , dpi ) pairs are evaluated for each point of the
sub-stroke, we generate a histogram comprising B bins. The
value of B decides the extent to which the angular values
are quantized. To start with, the votes corresponding to the
bins are initialised to zero. For each pi , we increment the
vote of the bin in which θpi falls into with a magnitude dpi .
In this manner, the accumulation of the votes across the bins
is done for all the N S points of S under consideration. The
resulting operation gives us a B dimensional vector which is
then normalized by its Euclidean norm to give hp .
Apart from the above, we generate two additional sets of
histograms that take into the account the information of the
vicinity around the sample points of a sub-stroke. Accordingly,
for each point pi in S, we employ a set of 2r + 1 points
{pi−r , ..., pi , ..., pi+r } in its neighbourhood Npi and compute
the following:

r
1
i = j pi+ j ∀i ∈ [r + 1, N S − r ]
2r + 1
j =−r
1  r
ai = j i+j ∀i ∈ [2r + 1, N S − 2r ] (2)
2r + 1
j =−r
The value of r determines the number of sample points in
Npi used for obtaining the above values at pi . In our work,
the value of r was set to two. This choice is reasonable,
considering that we do not have to lose more than four points
at the start and end of each stroke when compared to a higher
value of r .
Both i and ai , corresponding to point pi , encode the
change in the spatial coordinates and the rate of change
respectively. It may be mentioned that the form of Equation 2
bears semblance to the ‘delta’ and ‘acceleration’ coefficients
from the area of speech processing [37]. With regards to our
work, these consist of two components represented by i =
(i x , iy ) and ai = (ai x , aiy ) corresponding to the spatial x
and y coordinates respectively. Accordingly, we refer to them
as delta and acceleration vectors respectively. By noting that
the vectors encode the relative spatial information, we utilize
them to compute their orientation and magnitude as follows.
 
−1 iy
θi = tan
i x

di = (i x )2 + (iy )2 (3)
 
aiy
θai = tan−1
ai x

dai = (ai x )2 + (aiy )2 (4)
2542 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO. 10, OCTOBER 2018
Fig. 3. A toy-example for the computation of H B - (a) Illustration of the first level of k1 = 20 clusters obtained from the feature vectors of the three
writers with their prototypes and Voronoi cells highlighted. (b) Depiction of the three sub-clusters for a particular Voronoi cell shaded in yellow color in
sub-figure (a). In addition, we also present the computation of probability values { pik }3j =1 from the sub-cluster marked with an arrow.
Analogous to the generation of hp , a voting procedure for
the B bins of the histogram is applied on the (θi , di ) and
(θai , dai ) pairs to obtain h and ha respectively. Thereafter,
the vectors hp , h and ha are concatenated to provide the
feature vector description fS for S.
f S = [hp , h , ha ]
It can be noticed that the dimension D of fS is 3 × B. For
a document consisting of n T sub-strokes, we denote the set
of feature vectors as F, defined as:
F = {f1 , f2 , · · · , fn T } = {f j }nj =1
T
where f j is the 3 × B dimensional feature vector for the
j t h sub-stroke in the document.
A. Entropy Based Selection of Bin Size B
From the preceding discussion, it is evident that the dimen-
sion of the feature representation derived from the segmented
sub-strokes relies on the number of bins B used for generating
the set of histograms. In addition, from the work-flow of
Figure 1, it can be seen that these are utilized subsequently
to derive the writer descriptors. Hence, it becomes imperative
to select an appropriate value of B, so that increased discrim-
ination of descriptors between writers can be achieved. As a
step towards making this choice, we propose a scheme based
on analyzing the values of entropy obtained subsequent to a
two level clustering scheme as detailed below:
In the first level, feature vectors corresponding to the
sub-strokes are pooled among the set of paragraphs of the
W writers and partitioned into k1 clusters. The resulting
prototypes relate on an average to the shape characteristics
of the frequently occurring sub-strokes. In the second level,
we partition the feature vectors assigned to each of the k1
clusters separately into W clusters - corresponding to the
number of writers enrolled for training. For the sake of clarity,
we refer to each of the W clusters at the second level as a
‘sub-cluster’. Thus, in total, we obtain k1 × W sub-clusters at
the second level.
j
The fraction of feature vectors corresponding to each of the
W writers in a given sub-cluster can now be used to compute
the entropy measure as:

W
Hik =
j j
− pik log2 ( pik ) 1 ≤ i ≤ k1 , 1 ≤ k ≤ W (7)
j =1
(5)
From the preceding definition, the entropy values correspond-
ing to the W sub-clusters generated from the i t h cluster can
be written as {Hik }k=1
W
. Clearly, the calculation of a specific
Hik in Equation 7 involves obtaining the probability values
j
for the W writers, namely { pik }W j =1 .
(6) To motivate the use of entropy measure for the selection of
the bin size B, consider the scenario wherein the feature vec-
tors assigned to the k t h sub-cluster encapsulate the complete
information from only the n t h writer, 1 ≤ n ≤ W . Assuming
that the sub-clusters have been generated from the i t h cluster,
j
we can infer that the value of pik will be one for j = n and
zero otherwise. Accordingly from Equation 7, we see that the
entropy value Hik will achieve a value of zero.
As an extension of the above, when each of the k1 × W sub-
clusters generated in the second level comprise feature vectors
corresponding to only one writer, the average entropy measure
defined by
1 
k1 
W
H= Hik (8)
k1 × W
i=1 k=1
will again be zero. We therefore can infer that lower entropy is
a measure to represent better separability of the feature vectors
across the writers. Lastly, by recalling that the feature set of
the segmented sub-strokes themselves are parameterized with
a bin size B, we explicitly refer to the average entropy value
obtained in Equation 8 by H B . The choice of B (denoted
as B * ) that can be considered is as follows:
B * = arg min H B (9)
B
For a better understanding of the preceding discussion,
we provide a toy example illustration in Figure 3 with three
writers enrolled to the identification system. The feature vec-
tors of the writers are depicted with different symbols. In the
VENUGOPAL AND SUNDARAM: ONLINE WRITER IDENTIFICATION WITH SPARSE CODING-BASED DESCRIPTORS 2543

sub-figure (a), we depict the k1 clusters obtained at the first have n i ≤ n T . We define two scores S + −
pi (d) and S pi (d) for the
level with their corresponding Voronoi cells. In this example, t h t h
d attribute of the p feature vector f pi having a non-zero
the value of k1 used is 20. Thereafter, the feature vectors in a sparse coding coefficient α pi for the i t h dictionary atom φ i .
given Voronoi cell are used to generate W = 3 sub-clusters at ⎧
⎨ 1
the second level. The value of W corresponds to the number + f pi (d) ≥ α pi ×φi (d)
of writers enrolled in the system. The sub-clusters of one such S pi (d)= 1 + | f pi (d) − α pi × φi (d)|
⎩
Voronoi cell, highlighted with a yellow color in Figure 3(a) 0 other wi se
⎧
is shown in sub-figure (b). The entropy corresponding to the ⎨ 1
f pi (d) < α pi ×φi (d)
distribution of the writers in each of the three sub-clusters S− (d)= 1 + | f pi (d) − α pi × φi (d)|
pi ⎩
are computed separately and then accumulated. This process 0 other wi se
of accumulation is carried out across all the sub-clusters in 1 ≤ i ≤ K , 1 ≤ p ≤ n i (12)
the feature space, thus leading to the value of the average
entropy H B (refer Equation 8). For sake of completion, Here φi (d) is the value of the d t h attribute (1 ≤ d ≤ D) in
in the sub-figure 3(b), we also present the probability values the atom φ i . The Equation 12, in a sense, attempts to capture
corresponding to the writers of a sub-cluster (indicated with the error obtained along the d t h attribute while considering
an arrow). a single dictionary atom for reconstruction. If we define this
error as e pi (d) = f pi (d) − α pi × φi (d), then with regards to
V. S PARSE C ODING : A N OVERVIEW the attribute f pi (d) for which e pi (d) ≥ 0, we assign it with
The sparse framework aims at efficiently determining an the score S + pi (d). Likewise, when e pi (d) < 0, this attribute
over-complete set of basis vectors, in a way so that each feature is given a score of S − pi (d). In particular, we observe that the
vector can be approximated closely as a linear combination of more proximal f pi (d) is to α pi ×φi (d), the higher is the value
them. The term ‘sparse’ comes from the fact that most of the of the score S − +
pi (d) or S pi (d) (as applicable).
coefficients obtained while representing a feature vector are The obtained scores across the d t h feature attribute of
close to zero. In our work, we consider that a document with {f pi }np=1
i
are subsequently accumulated and normalized with
n T feature vectors {f j }nj =1 T
and an over-complete dictionary D respect to the values obtained from the entire document to
of K atoms, [φ φ 1 , φ 2 , ..., φ K ] are available to us. Any feature obtain S̃i+ (d) and S̃i− (d).
vector f j of a segmented sub-stroke can then be closely
ni
represented / approximated as a linear combination of the S+
pi (d)
atoms φ i , 1 ≤ i ≤ K with a set of coefficients α j = S̃i+ (d) =
p=1

[α j 1 α j 2 · · · α j K ]T as follows:
K
nj
S+
p j (d)

K j =1 p=1
fj ≈ α j iφ i (10) ni
S−
pi
i=1
p=1
Since K is higher than D (the dimension of f j ), there are S̃i− (d) = 1 ≤ i ≤ K, 1 ≤ d ≤ D (13)

K
nj
infinitely large solutions for Equation 10. In order to obtain S−
p j (d)
a unique set of coefficients, we impose a sparsity constraint j =1 p=1
that ensures that only a few of the K coefficients in α j are The scores for each of the D feature attributes are stacked to
non-zero. For a pre-learnt dictionary D of atoms, the values form the descriptors S+ −
i and Si as shown:
of these coefficients can be obtained by solving the following
+ T
optimization formulation: S+ +
i = S̃i (1) · · · S̃i (d) · · · S̃i+ (D)

− T

K S− −
i = S̃i (1) · · · S̃i (d) · · · S̃i− (D) (14)
min  f j − α j i φ i 2 +λ  α j 1 (11)
αj To ensure compactness of the descriptor for the i t h dictionary
i=1
atom, we consider the L 2 norm representation of the S+ i and
The factor λ is a Lagrange multiplier which controls the
S−
i vectors.
influence that the two terms have on each other. With regards
to the estimation / learning of the dictionary D in the training Si = [ S+ −
i 2 ,  Si 2 ]
T
(15)
phase, we use the algorithm of [38], implemented in the
SPArse Modeling Software (SPAMS) optimization toolbox. The final descriptor is obtained by concatenating the descrip-
tors across all the K dictionary atoms.6

T
VI. S PARSE C ODING BASED W RITER D ESCRIPTION S = S1 S2 · · · S K (16)
Prior to deriving the descriptor, the feature vectors {f j }nj =1
T

extracted from the n T sub-strokes, together with their respec-

tive sparse coefficient vectors {α α j }nj =1
T
are assumed to be
known. Let the subset of feature vectors in {f j }nj =1T
with a 6 We also explored on deriving descriptors by scoring the attributes without
consideration to the sign of the reconstruction error. However, the present
non-zero sparse coefficient α j i for the i dictionary atom
t h
strategy of using separate scores Si+ and Si− resulted in improved writer
ni
be represented as {f pi } p=1 . Without loss of generality, we identification rates. The details are presented in sub-section VIII-F.
2544 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO. 10, OCTOBER 2018

Algorithm 1 Proposed Writer Descriptor Generation

• Input:
– A set of histogram based feature vectors {f j }nj =1
T
extracted from a document with n T sub-strokes.
– Dictionary with K atoms {φ φ i }i=1 .
K

• Output:
– Writer descriptor S.
• Algorithm:
Compute:
◦ α j - the sparse coding coefficient vector for each feature vector f j .
◦ n i - the number of feature vectors having non-zero sparse coding coefficient for the i t h atom φ i .

% Descriptor Computation:
for every dictionary atom in {φ φ i }i=1
K do

for every feature vector having non-zero sparse-coding coefficient for φ i do

Compute S + −
pi (d) and S pi (d) scores across the D attributes
end for
end for

% Normalization Step:
for every dictionary atom in {φ φ i }i=1
K do

Calculate S̃i+ (d) and S̃i− (d) scores across the D attributes.
end for

Stack the S̃i+ (d) and S̃i− (d) scores across the D attributes to obtain S+ −
i and Si vectors respectively.
+ −
Evaluate L 2 norm of the Si and Si vectors to get Si .
Concatenate {Si }i=1
K to represent the writer descriptor S.

We see that the dimension of S is 2 × K . In the event obtained between the descriptors of two different writers
of a dictionary atom φ i not having any of the points in w1 , w2 as follows.
the document with a non-zero sparse coding coefficient, its M M
w1 w2
descriptor Si is set to the zero vector. To summarize our 1  
d2 (w1 , w2 ) =  Sw1 i − Sw2 j 1 (18)
strategy, the pseudo-code of our proposed descriptor is shown Mw1 × Mw2
i=1 j =1
in Algorithm 1.
Subsequent to the above calculations between each pair
A. Discussion
of writers, the average inter-writer City-block distance for
From Equation 15, it can be observed that we considered writer w1 can be written as:
stacking the L 2 norm of the S+ −
i and Si vectors for constructing
Si rather than the entire vectors themselves. The rationale 1 
W

behind this step is to construct a compact descriptor for a d2 (w1 ) = d2 (w1 , m) (19)
W −1
m=1
document in terms of feature dimension. If we consider the m =w1
entire S+ −
i and Si vectors for generating Si , our final descriptor
would have a dimension that is increased by a factor of D. The ratio R of the average inter-writer to the intra-writer
To ensure that the L 2 norm operation does maintain the distance is then defined by:
discrimination between the writers, we conduct an experiment
W
that determines the ratio of the average inter writer to intra d2 (i )
i=1
writer descriptor distance, as follows: R= (20)
Mw W
Let {Sw1 i }i=11 correspond to the final descriptors for the d1 (i )
Mw1 documents written by writer w1 . The average pair-wise i=1
intra-writer City Block distance between these descriptors is The value of R is related to the degree of separation of the
denoted as d1 (w1 ). writer descriptors in the feature space. In Table I, we present
Mw1 Mw1
1  the results on the data of the writers from the databases
d1 (w1 ) =  M
 Sw1 i − Sw1 j 1 (17) used in this experimentation, namely the IAM, IBM-UB1 and
w 1
2 i=1 j =1 CASIA.7 For this set-up, the value of bins for the histogram
j =i
based features and the dictionary size was set to 10 and 400,
As a first step to computing the inter writer descriptor
distance, we consider the average of the City-block distance 7 A detailed description of the databases are provided in Section VII-A.
VENUGOPAL AND SUNDARAM: ONLINE WRITER IDENTIFICATION WITH SPARSE CODING-BASED DESCRIPTORS
TABLE I
I LLUSTRATION OF THE R R ATIO AT PARAGRAPH AND T EXT L INE L EVEL
ON THE IAM, IBM-UB 1 AND CASIA D ATABASES . T HE F ULL
D ESCRIPTOR C ORRESPONDS TO U SING THE E NTIRE V ECTORS OF
S+ −
i AND Si A CROSS A LL THE D ICTIONARY ATOMS FOR
R EPRESENTATION . T HE C OMPACT D ESCRIPTOR , ON
THE O THER H AND , C ONSIDERS THE L 2 N ORM
respectively. The choice behind these values is based on the
discussion in subsection VIII-B.
We considered two possible representation of the descriptors
as outlined below:
+ −
• We stack the entire vector of Si and Si obtained across
all the K dictionary atoms, resulting in a 2 × D × K
dimensional descriptor. For sake of brevity, we refer to
this as the ‘full descriptor’.
+ −
• We stack the L 2 norm of the Si and Si vectors obtained
across all the K dictionary atoms. We refer to this as the
‘compact descriptor’.
With the above two representations, we obtain a comparable
value of the ratio R across all the databases at paragraph
and text-line levels respectively. This in turn suggests that
either of the descriptors may be used for writer identification.
Nonetheless, employing L 2 norm leads to a reduction in the
dimension by a factor D over the full descriptor.
VII. E XPERIMENTAL S ET-U P
In this section, we present an outline of the three databases
used together with their enrollment and evaluation protocols.
A. Database Description
The IAM Online Handwriting English database [39]:
comprises online handwriting data obtained on a white-board
from 217 writers. Each writer contributed to as much as eight
paragraphs of 50 words with the recordings of (x, y) spatial
coordinates and time stamps of the pen trace being noted.
The IBM-UB1 English database [40]: comprises samples
of 43 writers acquired from the IBM Cross-pad. For each topic
chosen by a specific writer, two documents are generated -
summary and ‘query’ text. The summary text contains a one
to two page essay on a particular topic while the query data
captures the essence of the summary text with approximately
25 words. The spatial x and y-coordinates are recorded for
each handwritten document in this database.
The CASIA Chinese database [41]: consists of texts
provided by the Institute of Automation, Chinese Academy of
Sciences. The online data being employed are contributed by
187 writers, with each writer providing three Chinese pages.
B. Training and Testing Protocol
While evaluating the performance of the proposal, two
protocols are considered with regards to the training and
testing of the handwritten documents. These are as follows.
2545
• For the experiments presented in subsections VIII-A to
VIII-F of this paper, the following protocol is suggested
for the IAM database. Four random paragraphs written by
each writer are selected for training while the remaining
are reserved for the evaluation of our proposal. The
protocol advocated in [19] is used for the IBM-UB 1
handwriting data-set with 80% of the randomly selected
summary document paragraphs from each writer used for
enrolment and the remaining for testing. For the CASIA
Chinese database, two random pages of text per user are
selected for enrolment. The performance of our algorithm
is evaluated on the remaining one page [33], [35].
For each database, the atoms of the dictionary are learnt
from the enrolled data corresponding to 25% of the
number of users, that are randomly chosen. The Support
Vector Machine (SVM) is used to establish the identity of
the writer for the test document. In particular, a one versus
all SVM with the Radial Basis Function (RBF) kernel is
considered. The optimal values of the parameters C and
γ are obtained via the grid search technique.
• In the second protocol, for each of the databases we
consider dividing the handwritten samples into a disjoint
training and test set, so that there is no overlap of
writers between them. The training set corresponds to
the data contributed from 25% of the number of users
that are randomly chosen. The test set comprises samples
of the remaining 75 % of writers. Moreover, in place
of the traditional one versus all SVM, the Exemplar-
SVM is used for writer identification [8]. The results
of our methodology using this set-up is presented in
Sub-section VIII-G.
We consider two types of training strategies for our experi-
mentation namely, paragraph and text-line level. In the for-
mer, the proposed descriptor is derived separately for each
paragraph while for the latter, we divide each paragraph
into individual text lines and construct the descriptor. While
evaluating the performance of our strategy, we mention the
average writer identification rate obtained from ten trials.
A trial comprises the following steps:- selection of documents
for enrolment (using one of the above protocols), dictionary
learning, construction of descriptor and classification.
VIII. R ESULTS AND D ISCUSSION
In this Section, we demonstrate several experiments that
reflect the various aspects of the proposal being explored for
establishing the authorship of online handwritten documents.
A. Analysis of Average Entropy Values H B With Bin Size B
We begin by investigating on the size of the bin B that can
be utilized for the histogram based feature sets while ensuring
discrimination between the writers. As part of the experimental
set up, the number of clusters k1 used at the first level is varied
from 25 to 100 in steps of 25. As discussed in Section IV-A,
each of these clusters are further split into W sub-clusters at
the second level for the computation of the average entropy
value H B . In our case, the values of W used are 217, 43 and
2546 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO. 10, OCTOBER 2018

TABLE II
VARIATION OF THE AVERAGE E NTROPY VALUES H B FOR VARYING N UMBER OF B INS B ON THE IAM, IBM-UB 1 AND CASIA D ATABASES . T HE VALUE
OF k1 D ENOTE THE N UMBER OF C LUSTERS AT THE F IRST L EVEL . F OR M ORE D ETAILS , P LEASE R EFER THE T EXT

TABLE III
C OMPARISON OF AVERAGE W RITER I DENTIFICATION R ATES ( IN %) AT THE PARAGRAPH AND T EXT L INE L EVEL FOR THE P ROPOSED D ESCRIPTOR
W ITH VARYING N UMBER OF B INS B AND D ICTIONARY S IZE K ON THE IAM D ATABASE . T HE B EST AVERAGE I DENTIFICATION R ATES
I S M ARKED IN B OLD . (a) PARAGRAPH L EVEL . (b) T EXT L INE L EVEL

187 - corresponding to the number of writers in the IAM,
IBM-UB1 and CASIA database respectively.
The average entropy H B is recorded for bin sizes B, ranging
from 2 to 16 in steps of 2 in Table II. We observe that its value
initially decreases with increasing values of B. Subsequent to
reaching a minimum, it begin to increase again for higher
number of bins. The lowest value of H B is obtained with the
bin size of 10 across all the databases.
As the number of bins increase, it is likely that for large
values, a perturbation to one or more of the orientations
θpi , θi , θai can lead to a change in the bin to which the votes
are assigned. This results in relatively higher variation between
the attributes of the histogram feature vectors corresponding
to the sub-strokes of the same writer - thereby leading to the
increase in the average entropy value. On the other hand,
the increased H B values at small number of bins can be
explained due to the generated feature vectors not being able
to adequately capture the nuances that discriminate the feature
vectors belonging to different writers.
B. Performance With Varying Bins B and Dictionary Size K
In this subsection, we evaluate the performance of the
proposed writer identification system with the variation of the
number of bins B for the histogram based features and the
number of the atoms K in the dictionary. We vary the value
of B from two to 16 in steps of two and number of atoms
from 50 to 500 in steps of 50. Table III (a) and III (b) report
the results of our strategy for the IAM database. We observe
that the proposed writer descriptor provides the best average
identification rate of 99.45 % and 90.28% at paragraph and text
line level respectively for a dictionary size K = 400 and bin
size B = 10. With regards to the IBM-UB 1 database (shown
in Table IV (a) and IV(b)), we achieve a best average writer
identification rate of 97.21% and 83.49% at paragraph and text
line level respectively for the same value of (B, K ). Lastly,
for the CASIA database presented in Table V (a) and V (b),
the best average paragraph and text-line writer identification
rate is reported to be 95.58% and 80.85% respectively.
It is interesting to note that the value of the bin size B
leading to the best average writer identification rate coincides
to that providing the lowest average entropy measure H B
in Table II on all the three databases. This validates the need to
select an appropriate choice of bin size for the extraction of the
histogram feature sets, thereby leading to better discrimination
of the sparse descriptors between the writers.
With regards to the trend in identification rates across all the
databases, for a fixed dictionary size as we vary the number of
VENUGOPAL AND SUNDARAM: ONLINE WRITER IDENTIFICATION WITH SPARSE CODING-BASED DESCRIPTORS 2547

TABLE IV
C OMPARISON OF AVERAGE W RITER I DENTIFICATION R ATES ( IN %) AT THE PARAGRAPH AND T EXT L INE L EVEL FOR THE P ROPOSED
D ESCRIPTOR W ITH VARYING N UMBER OF B INS B AND D ICTIONARY S IZE K ON THE IBM-UB1 D ATABASE . T HE B EST AVERAGE
I DENTIFICATION R ATE I S M ARKED IN B OLD . (a) PARAGRAPH L EVEL . (b) T EXT L INE L EVEL

TABLE V
C OMPARISON OF AVERAGE W RITER I DENTIFICATION R ATES ( IN %) AT THE PARAGRAPH AND T EXT L INE L EVEL FOR THE P ROPOSED
D ESCRIPTOR W ITH VARYING N UMBER OF B INS B AND D ICTIONARY S IZE K ON THE CASIA D ATABASE . T HE B EST AVERAGE
I DENTIFICATION R ATE I S M ARKED IN B OLD . (a) PARAGRAPH L EVEL . (b) T EXT L INE L EVEL

bins, the average identification rate increases. After reaching
a maximum value, it begins to decrease for larger values
of B. Moreover, it may be observed that the performance
on the IAM, IBM-UB1 and CASIA databases is inversely
related to values of H B reported in Table II. In other words,
lower identification rates for a dictionary size K correspond
to higher values in H B and vice-versa, thereby suggesting a
strong dependence.
Furthermore, it can also be noticed that for a particular value
of bin size B, the average writer identification rate increases
with dictionary size and then becomes comparable for values
of K ≥ 400. A small dictionary size leads to a low average
identification rate due to the limited prototypes available for
differentiating the sparse coding descriptors of the different
writers.
Considering that the best performance on all the data-
sets were obtained with writer descriptors constructed
from histogram features parameterized with B = 10,
we use this bin size for the experiments described in
subsections VIII-C, VIII-E, VIII-F and VIII-G.
C. Influence of the Segmentation Strategy
As mentioned in Section III, there were two strategies
described for the generation of sub-strokes. In this experiment,
we consider the performance of our proposal on each of them
separately. For sake of simplicity, we denote the method of
splitting the online trace at the local maxima in a stroke
as ‘SEG-1’. The second strategy ‘SEG-2’ corresponds to
generating sub-strokes with a fixed number of points (chosen
as 30 in this work).
Table VI presents the best average writer identification
results across all the databases at paragraph and text-line
levels, with the bin and dictionary size being specified. From
the entries of the third row, it may be inferred that the
substrokes from both the strategies SEG-1 and SEG-2 leads
to improvement of the writer identification system.
D. Influence of Histogram Feature Sets on Writer Description
In the preceding experiments, we evaluated the efficacy of
the writer descriptor constructed from the histogram based
2548 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO. 10, OCTOBER 2018

TABLE VI
P ERFORMANCE C OMPARISON OF O UR P ROPOSAL FOR THE D IFFERENT S EGMENTATION S TRATEGIES .
T HE N UMBERS B EING M ENTIONED A RE THE I DENTIFICATION R ATES ( IN %)

TABLE VII
P ERFORMANCE C OMPARISON OF P ROPOSED W RITER I DENTIFICATION S YSTEM ( IN %) W ITH D IFFERENT C OMBINATIONS OF H ISTOGRAM F EATURE S ETS

feature sets (hp , h and ha ) extracted from the sub-strokes TABLE VIII
(discussed in Section IV). We now validate the influence C OMPARISON OF AVERAGE W RITER I DENTIFICATION R ATES ( IN %) AT
THE PARAGRAPH L EVEL FOR THE D ESCRIPTOR O BTAINED VIA
of the different choice of feature sets for writer description k-M EANS AND T HROUGH S PARSE C ODING W ITH VARYING
in Table VII. The best writer identification accuracies are D ICTIONARY S IZES K . T HE B EST AVERAGE
reported with the number of dictionary atoms K and bin I DENTIFICATION R ATES A RE M ARKED
IN B OLD FOR E ACH D ATABASE
size B being specified. The bin size for each combination
of histogram sets corresponds to the one minimizing the
average entropy value H B (defined in sub-section IV-A). From
the entries of the last row, we see that the augmentation of
feature sets h and ha (that capture the vicinity information)
improve the performance of the identification system beyond
that provided by hp .

E. Impact of the Sparse Framework for Writer Description

As mentioned in Section I, sparse coding framework offers
flexibility in the representation of the segmented sub-strokes
over hard clustering algorithms such as k-means. To demon-
strate this, we investigate its influence on writer description
over the k-means algorithm for varying dictionary sizes/
number of prototypes. The average writer identification rates
obtained at the paragraph level for the IAM, IBM-UB1 and
CASIA databases are shown in Table VIII. The sparse-based
descriptors are derived from the histogram feature set com-
bination (hp , h and ha ) extracted from the segmented sub- the sparse coding descriptors achieve an average identification
strokes of the handwritten data. It may be noted that across rate of 99.45% for K = 400. This, as we see, corre-
the different dictionary sizes considered, they provide a better sponds to a performance improvement of 2.73%. For the
performance to those derived via k-means for all the three IBM-UB1 database, a best average writer identification rate
databases. of 97.21 % is achieved with the descriptors derived via
With regards to the IAM database, the descriptors from sparse coding for a dictionary size K = 400, which is an
the codebook generated with the k-means attain a best improvement of 3.59% over those obtained from the k-means.
performance of 96.81% for K = 450. Contrast to this, For the CASIA database the best result of 95.58% is obtained
VENUGOPAL AND SUNDARAM: ONLINE WRITER IDENTIFICATION WITH SPARSE CODING-BASED DESCRIPTORS
Fig. 4. Variation of average writer identification rates (in %) at the text-line level for the proposed descriptor obtained via k-means (shown in blue) and
through sparse coding (in red) on the IAM (sub-figure (a)), IBM-UB1 (sub-figure (b)) and CASIA database (sub-figure (c)).
using the sparse framework, again with K = 400. This value
outperforms that of 88.43% for the k-means.
Moving further, an experimental comparison was also per-
formed at the text line level across all the databases. The
average writer identification rates for varying dictionary sizes
are presented in Figure 4. Note, that for sake of simplicity,
we abbreviate it as IR. From the best average writer identifica-
tion results presented in the sub-figures, the descriptors from a
sparse coding framework provide an improvement of 19.02%,
6.98% and 23.21% for the IAM, IBM-UB 1 and CASIA
databases respectively.
F. Performance Variation With a Variant of Writer Descriptor
Recall that in the derivation of the writer descriptor in
Section VI, we separately score each attribute of the seg-
mented sub-stroke by taking into consideration whether the
reconstruction error is positive or negative. For the sake of
completion, we also explored the possibility of a variant
of the descriptor, wherein for a dictionary atom, attributes
are scored irrespective of the sign of the reconstruction
error. Mathematically speaking, the score of the d t h attribute
(1 ≤ d ≤ D) of the pt h feature vector f pi having a non-zero
sparse coding coefficient α pi for the i t h dictionary atom φ i is
given as:
1 corresponding values of bin size and dictionary size (B, K ) are
S pi (d) =
1 + | f pi (d) − α pi × φi (d)|
1 ≤ i ≤ K , 1 ≤ p ≤ n i (21)
Subsequent to obtaining the scores, the descriptor is computed
by following steps similar to those described from Equation 13
to Equation 15. However, instead of obtaining two separate
vectors S+ −
i and Si for the i
t h dictionary atom φ , we get
i
a single vector of dimension D in this variant. The final
descriptor then comprises the L 2 norm of these vectors,
computed across all the K dictionary atoms - thus leading
to a writer descriptor of dimension K . The histogram feature
set combination (hp ,h and ha ) extracted from the segmented
substrokes are used for this experiment.
The evaluation of the above descriptor at the paragraph
and text line level is presented in Table IX. For sake of
brevity, we refer it as the ‘K -dim’, while the one discussed
in Section VI is abbreviated to as ‘2 × K -dim’. The best
average writer identification rate along with the dictionary
2549
size K is mentioned in the Table. From the entries, we can
infer that across the data-sets, the ‘2 × K -dim’ descriptor
provides improved results at paragraph and text-line levels,
when compared to its variant - ‘K -dim’. This, we believe,
may be due to the increased discrimination achieved with the
scoring of the attributes of the segmented sub-stroke feature
vectors based on both the value and sign of the reconstruction
error. Needless to say, the higher performance is achieved with
the trade-off of increasing the dimension by a factor of nearly
two (450 versus 800).
G. Performance Evaluation Using Exemplar-SVM on a
Disjoint Train and Test Dataset
In this subsection, we evaluate the efficacy of our proposal
using a different training and test protocol. We divide the
handwritten samples of a database into a disjoint training
and test set, so that there is no overlap of writers between
them. The training set corresponds to the data contributed
from 25% of the number of users that are randomly chosen
in each trial. The test set comprises samples of the remaining
75 % of writers. Moreover, for establishing the authorship of
a document, we explore the utility of the Exemplar-SVM
framework8 recently suggested by Christlein et al. [8].
The best average writer identification results are presented
in the Table X for the different sets of histogram features. The
also mentioned for each combination. Apart from this, we also
evaluated our proposal with the different segmentation strate-
gies using the Exemplar-SVM. The results for this experiment
are given in Table XI.
IX. D ISCUSSION TO P RIOR W ORKS
In this Section, we see how the performance of the proposed
writer identification system fares to the previous explorations
that use the IAM, IBM-UB 1 and CASIA databases (refer
Table XII). Since our method relies on hand-crafted features,
we begin the discussion by enumerating all prior systems from
literature in this paradigm. Unless specifically mentioned as a
table-note, the entries of the Table are the writer identification
rates (in %) as reported from the respective references. It may
8 In the traditional SVM formulation, given a new handwriting sample from
an unseen writer, all models need to be retrained. This issue is however
alleviated by the use of the Exemplar-SVM.
2550 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO. 10, OCTOBER 2018

TABLE IX
P ERFORMANCE C OMPARISON W ITH A VARIANT OF THE W RITER D ESCRIPTOR ON THE T HREE D ATABASES . T HE N UMBERS B EING M ENTIONED A RE
THE I DENTIFICATION R ATES ( IN %)

TABLE X
P ERFORMANCE C OMPARISON OF P ROPOSED W RITER I DENTIFICATION S YSTEM ( IN %) W ITH D IFFERENT C OMBINATIONS OF H ISTOGRAM F EATURE
S ETS . T HE T RAINING AND T ESTING P ROTOCOL OF S UBSECTION VIII-G I S U SED

TABLE XI
P ERFORMANCE C OMPARISON OF O UR P ROPOSAL FOR THE D IFFERENT S EGMENTATION S TRATEGIES U SING THE T RAINING AND T ESTING P ROTOCOL OF
S UBSECTION VIII-G. T HE N UMBERS B EING M ENTIONED A RE THE I DENTIFICATION R ATES ( IN %)

also be noted that the systems being outlined can differ with
regards to the features as well as the classifier and enrolment
strategies.
To the best of our knowledge, the writer identification
systems proposed in [12], [13], [19], [21], and [22] utilize
the IAM database for experimentation. Schlapbach et al. [13]
use a GMM based framework and present an identification
rate of 98.56% and 88.96% at paragraph and text line level.
The concept of Latent Dirichlet Allocation utilized in [19]
and the subtractive clustering based approach of [21] lead to a
paragraph-level writer identification rate of 93.39% and 96.3%
respectively. The incorporation of the tf-idf framework on the
sparse representation framework in [12] achieved a paragraph
level identification rate of 98.94% and a text line level accu-
racy of 83.30%. An attempt on writer description from a pre-
generated code-book using the k-means algorithm resulted in
a performance of 97.81% and 80.61% at paragraph and text-
line level [22]. In addition, we wish to state that the results of
98.54% and 82.18% reported for the work [6] correspond to
our implementation with the online handwritten data9 (using
the feature vectors of the segmented sub-strokes described
in Section IV). On the whole, the proposed methodology
attains higher average writer identification rates of 99.45% and
90.28% at paragraph and text-line level using the traditional
SVM. The Exemplar-SVM also presents competing results
of 99.40% and 89.88%.
To the best of our knowledge, only two studies have
been made to date on the IBM-UB1 database [19], [22].
These works provide an average writer identification accu-
racy of 89.47% and 94.37% at the paragraph level. Con-
trast to these and our implementation of the scheme [6],
the proposed methodology of sparse coding based descrip-
tors does remarkably better with both traditional and
Exemplar-SVM.
With respect to the CASIA Chinese database, we see that
Li et al. [27] mention Top 1 writer identification rate of 80%
9 Kumar et al. [6] evaluated their proposal on off-line handwritten data.
VENUGOPAL AND SUNDARAM: ONLINE WRITER IDENTIFICATION WITH SPARSE CODING-BASED DESCRIPTORS 2551

TABLE XII
S URVEY OF O NLINE W RITER I DENTIFICATION S YSTEMS ON THE IAM, IBM-UB 1 AND CASIA D ATABASES . U NLESS S PECIFICALLY
M ENTIONED AS A TABLE -N OTE , THE N UMBERS A RE THE W RITER I DENTIFICATION R ATES ( IN %) AS R EPORTED F ROM THE
R ESPECTIVE R EFERENCES . M OREOVER , E NTRIES IN THE TABLE M ARKED W ITH (-) S YMBOL I NDICATE T HAT THE
R ESPECTIVE AUTHORS H AVE N OT R EPORTED THE R ESULTS FOR T HEM

at paragraph-level for the GMM-UBM framework [13]. X. C ONCLUSION

Moreover, the hierarchical shape primitive approach by the
same authors report an accuracy of 91.20%. Our proposal out-
performs not only these methods but also the implementation
of the sparse representation framework [6] with our histogram-
based features.
It may thus be inferred that amongst the several state-of-
art methods presented in Table XII with hand-crafted features,
our algorithm achieves a better performance across all the
databases.
Lastly, there have been recent developments in online writer
identification with deep learning techniques. These methods
learn the feature representation from the handwritten data,
thus alleviating the need of hand-crafted features. In particular,
a Convolutional Neural Network based architecture on the
CASIA Chinese database provides a writer identification rate
of 95.72 % [33]. This is almost at par to the best result
of 95.58 % obtained with our proposal at paragraph-level.
The most recent work on the same database (published in
April 2017), however, uses Recurrent Neural Networks to
achieve a performance of up-to 99.5% [35].
Based on the preceding discussion, we see that our method
is quite promising for identifying the authorship of online
handwritten text documents. The obtained results are better
than several state-of-art works.
In this work, we presented an online writer identification
system by employing a sparse framework. Our contributions
are enumerated as follows: (i) Proposal of descriptors derived
from a set of dictionary atoms. The descriptors for each
dictionary atom encode the error while using it alone for
reconstruction. (ii) Utilization of histogram based feature
vectors extracted at a sub-stroke level, together with their cor-
responding sparse coefficients, for deriving the sparse coded
descriptors. (iii) Proposal of an entropy based analysis for the
appropriate bin size to be selected for obtaining the features
so as to ensure discrimination between the writer descriptors.
(iv) Evaluation of the strategy on a number of databases with
results demonstrated to be on par or even better to several
prior works.
We conclude this paper with a terse discussion on how our
proposal can be extended for the offline writer identification
setting. The road-map of steps would be to first segment
out the allographs or letter parts and then describe each
patch thereof with features such as SIFT or RootSIFT [8].
The extracted features can be utilized to construct the over-
complete dictionary and subsequently the writer descriptor,
as outlined in Section VI. At this juncture, it may be worth
mentioning that, in place of hand crafted SIFT based features,
2552 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO. 10, OCTOBER 2018
one can also attempt to describe the allographs with con-
volutional neural network (CNN) activation features as done
in [42] and [43].
R EFERENCES
[1] A. K. Jain, A. Ross, and S. Prabhakar, “An introduction to biometric
recognition,” IEEE Trans. Circuits Syst. Video Technol., vol. 14, no. 1,
pp. 4–20, Jan. 2004.
[2] A. A. Brink, J. Smit, M. L. Bulacu, and L. R. B. Schomaker, “Writer
identification using directional ink-trace width measurements,” Pattern
Recognit., vol. 45, no. 1, pp. 162–171, 2012.
[3] D. Bertolini, L. S. Oliveira, E. Justino, and R. Sabourin, “Texture-based
descriptors for writer identification and verification,” Expert Syst. Appl.,
vol. 40, no. 6, pp. 2069–2080, 2013.
[4] S. He and L. Schomaker, “Co-occurrence features for writer identifica-
tion,” in Proc. 15th Int. Conf. Frontiers Handwriting Recognit. (ICFHR),
Shenzhen, China, Oct. 2016, pp. 78–83.
[5] J. Wen, B. Fang, J. Chen, Y. Tang, and H. Chen, “Fragmented edge
structure coding for Chinese writer identification,” Neurocomputing,
vol. 86, pp. 45–51, Jun. 2012.
[6] R. Kumar, B. Chanda, and J. Sharma, “A novel sparse model
based forensic writer identification,” Pattern Recognit. Lett., vol. 35,
pp. 105–112, Jan. 2014.
[7] X. Wu, Y. Tang, and W. Bu, “Offline text-independent writer identi-
fication based on scale invariant feature transform,” IEEE Trans. Inf.
Forensics Security, vol. 9, no. 3, pp. 526–536, Mar. 2014.
[8] V. Christlein, D. Bernecker, F. Hönig, A. Maier, and E. Angelopoulou,
“Writer identification using GMM supervectors and exemplar-SVMs,”
Pattern Recognit., vol. 63, pp. 258–267, Mar. 2016.
[9] J. Wang, C. Lu, M. Wang, P. Li, S. Yan, and X. Hu, “Robust face
recognition via adaptive sparse representation,” IEEE Trans. Cybern.,
vol. 44, no. 12, pp. 2368–2378, Dec. 2014.
[10] Y. Gao, J. Ma, and A. L. Yuille, “Semi-supervised sparse representa-
tion based classification for face recognition with insufficient labeled
samples,” IEEE Trans. Image Process., vol. 26, no. 5, pp. 2545–2560,
May 2017.
[11] Y. Liu, Z. Yang, and L. Yang, “Online signature verification based on
DCT and sparse representation,” IEEE Trans. Cybern., vol. 45, no. 11,
pp. 2498–2511, Dec. 2014.
[12] I. Dwivedi, S. Gupta, V. Venugopal, and S. Sundaram, “Online writer
identification using sparse coding and histogram based descriptors,”
in Proc. 15th Int. Conf. Frontiers Handwriting Recognit., Oct. 2016,
pp. 572–577.
[13] A. Schlapbach, M. Liwicki, and H. Bunke, “A writer identification
system for on-line whiteboard data,” Pattern Recognit., vol. 41, no. 7,
pp. 2381–2397, 2008.
[14] M. Liwicki, A. Schlapbach, and H. Bunke, “Fusing asynchronous
feature streams for on-line writer identification,” in Proc. 9th Int. Conf.
Document Anal. Recognit. (ICDAR), Sep. 2007, pp. 103–107.
[15] G. X. Tan, C. Viard-Gaudin, and A. C. Kot, “A stochastic nearest
neighbor character prototype approach for online writer identification,”
in Proc. 19th Int. Conf. Pattern Recognit., Dec. 2008, pp. 1–4.
[16] G. Tan, C. Viard-Gaudin, and A. Kot, “Online writer identification using
fuzzy c-means clustering of character prototypes,” in Proc. ICFHR,
Aug. 2008, pp. 475–480.
[17] C. Viard-Gaudin, G. X. Tan, and A. C. Kot, “Individuality of alphabet
knowledge in online writer identification,” Int. J. Document Anal.
Recognit., vol. 13, no. 2, pp. 147–157, 2010.
[18] A. Shivram, C. Ramaiah, U. Porwal, and V. Govindaraju, “Modeling
writing styles for online writer identification: A hierarchical Bayesian
approach,” in Proc. Int. Conf. Frontiers Handwriting Recognit. (ICFHR),
Sep. 2012, pp. 387–392.
[19] A. Shivram, C. Ramaiah, and V. Govindaraju, “A hierarchical Bayesian
approach to online writer identification,” IET Biometrics, vol. 2, no. 4,
pp. 191–198, 2013.
[20] C. Ramaiah, A. Shivram, and V. Govindaraju, “Data sufficiency for
online writer identification: A comparative study of writer-style space
vs feature space models,” in Proc. Int. Conf. Pattern Recognit. (ICPR),
Aug. 2014, pp. 3121–3125.
[21] G. Singh and S. Sundaram, “A subtractive clustering scheme for text-
independent online writer identification,” in Proc. 13th Int. Conf. Doc-
ument Anal. Recognit. (ICDAR), Aug. 2015, pp. 311–315.
[22] V. Venugopal and S. Sundaram, “An online writer identification system
using regression-based feature normalization and codebook descriptors,”
Expert Syst. Appl., vol. 72, pp. 196–206, Apr. 2017.
[23] H. Jégou, F. Perronnin, M. Douze, J. Sánchez, P. Pérez, and
C. Schmid, “Aggregating local image descriptors into compact codes,”
IEEE Trans. Pattern Anal. Mach. Intell., vol. 34, no. 9, pp. 1704–1716,
Sep. 2012.
[24] A. Namboodiri and S. Gupta, “Text independent writer identification
from online handwriting,” in Proc. 10th Int. Workshop Frontiers Hand-
writing Recognit., Oct. 2006, pp. 23–26.
[25] Z. Sun, B. Li, and T. Tan, “Online text-independent writer identification
based on stroke’s probability distribution function,” in Proc. Int. Conf.
Biometrics, Aug. 2007, pp. 201–210.
[26] B. Li and T. Tan, “Online text-independent writer identification based
on temporal sequence and shape codes,” in Proc. IEEE 10th Int. Conf.
Document Anal. Recognit. (ICDAR), Jul. 2009, pp. 931–935.
[27] B. Li, Z. Sun, and T. Tan, “Hierarchical shape primitive fea-
tures for online text-independent writer identification,” in Proc. IEEE
10th Int. Conf. Document Anal. Recognit., (ICDAR). Jul. 2009,
pp. 986–990.
[28] A. Chaabouni, H. Boubaker, M. Kherallah, A. M. Alimi, and H. El Abed,
“Multi-fractal modeling for on-line text-independent writer identifi-
cation,” in Proc. Int. Conf. Document Anal. Recognit., Sep. 2011,
pp. 623–627.
[29] M. Gargouri, S. Kanoun, and J.-M. Ogier, “Text-independent writer
identification on online Arabic handwriting,” in Proc. IEEE 10th Int.
Conf. Document Anal. Recognit., (ICDAR), Aug. 2013, pp. 428–432.
[30] T. Dhieb, W. Ouarda, H. Boubaker, M. B. Halima, and A. M. Alimi,
“Online Arabic writer identification based on beta-elliptic model,”
in Proc. Int. Conf. Intell. Syst. Design Appl. (ISDA), Dec. 2015,
pp. 74–79.
[31] T. Dhieb, W. Ouarda, H. Boubaker, and A. M. Alimi, “Beta-elliptic
model for online writer identification from online Arabic handwriting,”
J. Inf. Assurance Security, vol. 11, no. 5, pp. 263–272, 2016.
[32] W. Ouarda, T. Dhieb, H. Boubaker, and A. M. Alimi, “Deep neural
network for online writer identification using beta-elliptic model,” in
Proc. Int. Joint Conf. Neural Netw., Jul. 2016, pp. 1863–1870.
[33] W. Yang, L. Jin, and M. Liu, “DeepwriterID: An end-to-end online text-
independent writer identification system,” IEEE Intell. Syst., vol. 31,
no. 2, pp. 45–53, Mar. 2016.
[34] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and
R. Salakhutdinov, “Dropout: A simple way to prevent neural networks
from overfitting,” J. Mach. Learn. Res., vol. 15, no. 1, pp. 1929–1958,
2014.
[35] X.-Y. Zhang, G.-S. Xie, C.-L. Liu, and Y. Bengio, “End-to-end
online writer identification with recurrent neural network,” IEEE Trans.
Human-Mach. Syst., vol. 47, no. 2, pp. 285–292, Apr. 2016.
[36] N. Dalal and B. Triggs, “Histograms of oriented gradients for human
detection,” in Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern
Recognit., vol. 1. Jun. 2005, pp. 886–893.
[37] S. Furui, “Speaker-independent isolated word recognition based on
emphasized spectral dynamics,” in Proc. IEEE Int. Conf. Acoust.,
Speech, Signal Process. (ICASSP), vol. 11. Apr. 1986, pp. 1991–1994.
[38] J. Mairal, F. Bach, J. Ponce, and G. Sapiro, “Online learning for
matrix factorization and sparse coding,” J. Mach. Learn. Res., vol. 11,
pp. 19–60, Mar. 2010.
[39] M. Liwicki and H. Bunke, “IAM-OnDB—An on-line English sen-
tence database acquired from handwritten text on a whiteboard,”
in Proc. 8th Int. Conf. Document Anal. Recognit., Sep. 2005,
pp. 956–961.
[40] A. Shivram, C. Ramaiah, S. Setlur, and V. Govindaraju, “IBM_UB_1:
A dual mode unconstrained English handwriting dataset,” in Proc. 10th
Int. Conf. Document Anal. Recognit. (ICDAR), Aug. 2013, pp. 13–17.
[41] The Institute of Automation, Chinese Academy of Sciences (CASIA).
(2007). (CASIA) Handwriting Database—Biometrics Ideal Test.
[Online]. Available: http://biometrics.idealtest.org/
[42] V. Christlein, D. Bernecker, A. K. Maier, and E. Angelopoulou,
“Offline writer identification using convolutional neural network acti-
vation features,” in Proc. German Conf. Pattern Recognit., Nov. 2015,
pp. 540–552.
[43] Y. Tang and X. Wu, “Text-independent writer identification via CNN
features and joint Bayesian,” in Proc. 15th Int. Conf. Frontiers Hand-
writing Recognit., (ICFHR), Shenzhen, China, Oct. 2016, pp. 566–571.
Vivek Venugopal, photograph and biography not available at the time of
publication.
Suresh Sundaram, photograph and biography not available at the time of
publication.