Neural Computing and Applications (2023) 35:7509–7527

https://doi.org/10.1007/s00521-022-08045-z(0123456789().,-volV)(0123456789().,-volV)

ORIGINAL ARTICLE

On the performance analysis of various features and classifiers

for handwritten devanagari word recognition
Sukhjinder Singh1,4 · Naresh Kumar Garg2,4 · Munish Kumar3

Received: 27 April 2022 / Accepted: 7 November 2022 / Published online: 23 November 2022
© The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2022

Abstract
Holistic-based or segmentation-free handwritten word recognition is one of a pattern recognition problem that aims to
recognize the entire word image as a single entity. It is a new modality that recognizes handwritten words from its overall
shape and performs better than its complement, i.e., analytic approach for given small lexicon size. Due to technological
advancements, society is becoming paperless and prefers to use digital platform for various tasks. This paper deals with the
use of holistic approach for the recognition of offline handwritten Devanagari words based on some statistical features.
Feature vector sets have been generated to describe each word in the feature space by extracting unıform zoning-, diagonal-
and centroid-based features from the database of handwritten word images (50-word classes). Various classifiers, namely
K-nearest neighbor (KNN), decision tree and random forest, are employed for the recognition purpose. Furthermore to
enhance the system performance, combination of above mentioned features along with gradient boosted decision tree
algorithm is proposed. In this way, proposed system achieved maximum recognition accuracy of 94.53% and the attained
competent results are comparable with exiting state-of-the-art methods. Moreover, the proposed system has achieved F1-
score of 94.56%, FAR of 0.11%, FRR of 5.46%, MCC of 0.945 and AUC of 97.21%.

Keywords Handwritten word recognition · Holistic approach · Feature extraction · Classification · Gradient boosted decision
tree

1 Introduction

Handwritten word recognition (HWR) is an interesting and

& Sukhjinder Singh challenging research problem related with pattern recogni-
sukhjinder.ece@mrsptu.ac.in tion, artificial intelligence and machine vision [20, 24]. It
Naresh Kumar Garg becomes more challenging because of variation in writing
naresh2834@rediffmail.com style of writers due to their gender difference, age, qualifi-
Munish Kumar cation, occupation and temper [17], Roy et al. 28). It finds
munishcse@gmail.com many real-world applications including cheque reading,
1
editing old documents, handwritten notes reading, postal
Department of Electronics and Communication Engineering, automation, signature verification, meaning translation,
Maharaja Ranjit Singh Punjab Technical University, Bathinda,
Punjab 151001, India writer verification, reading aid for blind and visually
2 impaired documents and so on [5, 8, 12, 14, 28, 33, 36].
Department of Computer Science and Engineering, Maharaja
Ranjit Singh Punjab Technical University, Bathinda, Handwritten word recognition is an ability of a computer to
Punjab 151001, India receive and recognize handwritten words taken from vari-
3
Department of Computational Sciences, Maharaja Ranjit ous sources such as paper documents or touch screen,
Singh Punjab Technical University, Bathinda, Punjab 151001, written in any natural script such as Latin, Devanagari and
India Gurumukhi [11]. Devanagari script consists of 47 primary
4
Giani Zail Singh Campus College of Engineering and characters (14 vowels and 33 consonants) and is being used
Technology, Maharaja Ranjit Singh Punjab Technical for around 120 languages [7]. It is the fifth most well-
University, Bathinda, Punjab 151001, India

123
7510
known Indian script in the world [4]. Generally, handwritten
word recognition can be classified in two modes, namely
online and offline handwritten word recognition modes.
Online-HWR-based systems deal with the automatic con-
version or recognition of words as it is written on special
digitizer surface such as tablet and mobile phone using
special stylus pen [18]. In such systems, words shall be
recognized at real-time as soon as these are written on
digitizer surface. On the other hand, offline-HWR based
systems deal with the automatic conversion or recognition
of words from scanned images or paper documents by
capturing the information from these images or documents
[11]. Development of offline-HWR based systems becomes
more challenging due to non-availability of information
regarding stokes, speed and up/down movement of pen
[37]. Furthermore, approaches for the recognition of hand-
written words can broadly be classified into two categories,
namely analytical approach (or segmentation based) and
holistic approach (or segmentation-free) [6, 18]. In an
analytical approach, word is considered as the collection of
individual characters so it is segmented into individual units
or characters before recognition. Segmentation becomes
more challenging, if overlapping characters exist in the
word. Therefore to overcome such issue, holistic approach
can be used that considers the whole word as a separate
entity. So, there is no issue of explicit segmentation. The
whole word is recognized from its shape without segmen-
tation into separate words. From last few years, holistic
approach is being used for word recognition by various
researchers working in this area as it gives better recogni-
tion results as compared with analytical approach [37]. So
in this work, holistic approach has been used to recognize
handwritten Devanagari words as it also avoids the over-
head of segmentation. The presented work has been eval-
uated on a collected database of Devanagari words on the
basis of various performance metrics.
1.1 Motivations
The motivational factor behind this work is that holistic
approach based algorithms are mathematical competent for
small fixed-sized lexicon. After critically analyzing the lit-
erature, it has been gathered that the problems associated
with analytical approaches are due to various factors
including overlapping & touching characters and poor &
cursive writing [8, 10, 36, 37]. Most of the proposed
research works do not offer competent solution for the
same. So, in this work, holistic based offline-HWR system
for Devanagari script has been presented and analyzed in
terms of various performance metrics.
123
Neural Computing and Applications (2023) 35:7509–7527
1.2 Significance and contributions of the present
work
Literature available for holistic-based offline handwritten
Devanagari word recognition is very less. So authors have
implemented holistic approach for recognition of offline
handwritten Devanagari words by considering three state-
of-the-art features (uniform zoning-, diagonal- and centroid-
based) and classifiers (K-nearest neighbor, decision tree and
random forest). To improve the system performance, com-
bination of the above-mentioned features and gradient
boosted decision tree have also been used in this work. For
this work, authors have collected a dataset of 20,000
handwritten Devanagari words (samples) having 50-word
classes (city/place names) from hundreds of writers from
diverse background. Also, the performance of proposed
work has been compared with some other existing methods
and syntactic analysis is presented. This work finds its
application for postal automation.
The rest of the paper is organized as follows: Sect. 2
gives related work. General framework for HWR is pre-
sented in Sect. 3 along with the details of collected database
and methods considered for feature extraction and classifi-
cation for this work. Experiment results are presented and
discussed in Sect. 4 along with performance metrics and
comparative analysis. Finally, conclusions and future scope
are drawn in Sect. 6.
2 Related work
This section summarizes the research work or contribution
by various researchers working in the field of handwritten
word recognition system. Parui and Shaw [25] recognized
Devanagari words using stroke based features and Hidden
Markov Model (HMM) which uses statistical as well as
structural information of handwritten samples. Authors
applied some preprocessing steps on collected database so
as to minimize the writing variability effect due to different
writing styles of writers. They achieved promising results
for dataset with small lexicon size. Shaw et al. [33] devel-
oped an offline Devanagari handwritten word recognition
system based on directional chain code and HMM. Authors
used a database of 22,500 (training) & 17,200 (testing)
words and achieved 80.2% accuracy in their work. Fur-
thermore, in [34], authors extracted stroke based features
and used HMM classification approach using same dataset
and obtained 84.31% accuracy. Shaw and Parui [30]
developed a system for handwritten Devanagari words
taking 50 different city names are as database. In first stage,
authors extracted shape/stroke features and used HMM
classification approach. They obtained 82.89% and 87.71%
Neural Computing and Applications (2023) 35:7509–7527
recognition rates (first stage) on validation set and training
set, respectively, whereas, in second stage wavelet trans-
form based features were extracted and used modified
Bayes discriminant function for classification. They
achieved 85.57% and 91.25% test and training accuracies
(second stage). Singh et al. [35] developed a handwritten
word recognition system considering curvelet transform,
SVM and KNN taking a dataset of 28,500 words and
obtained encouraging results. They gathered that KNN gave
better results than the SVM with curvelet features and
obtained 93.21% accuracy for their dataset. Ramachandrula
et al. [27] proposed a handwritten recognition system for
Hindi words, by considering some approaches based on
directional element features and dynamic programming.
They achieved accuracies between 79.94% (30 vocabulary
words) and 91.23% (10 vocabulary words) on their col-
lected database from 100 writers.
Shaw et al. [32] developed an offline handwritten word
recognition technique for Devanagari script by formulating
directional distance distribution (DDD) and gradient-struc-
tural concavity (GSC)-based features using 100 city names
and achieved better recognition accuracy. They analyzed
combinations of DDD and GSC features and used SVM
classifier for word recognition. Authors concluded that
combination of SVM outputs of DDD and GSC features
gives improved recognition accuracy. Kumar [14] devel-
oped a recognition system for isolated Devanagari words
(hand printed)-based segmentation approach using a data-
base of 3500 words. They considered multi-layer perceptron
(MLP) algorithm for classification. Authors used separate
classifiers for the characters of upper, lower and middle
regions. Authors obtained maximum 80.8% accuracy using
neighbor pixel weight and gradient features and MLP
classification. Malakar et al. [20] developed a word recog-
nition system based on segmentation-free approach for
Hindi text. In their work, they extracted various features,
namely area, aspect ratio, density, pixel ratio, longest run,
centroid and projection length from word images. They
constructed a feature vector consisting of 89-elements and
used MLP-based classification along with some other
approaches. They obtained 96.82% success rate using MLP
classifier (sevenfold). Bhunia et al. [1] developed a novel
method for cross-language handwritten text recognition and
word spotting, for Bangla, Devanagari and Gurumukhi
scripts. Authors preferred zone-wise character mapping so
that efficient learning can be achieved. The working of their
framework depends upon the similarity between the input
and target scripts. They also used entropy-dependent script
similarity score by considering the mapping between input
and target characters. The disadvantage of their framework
is that performance of their proposed system depends upon
how the zones are segmented. Improper segmented zones
may result poor performance.
7511
Ghosh et al. [5] developed a system for recognition of
handwritten Bangla word images using segmentation-free
method. With the aim of improving the classification and
dimensionality reduction feature vector, they used wrapper-
filter approach based on Memetic Algorithm (MA). Authors
considered grid-based gradient and statistical & contour-
based features (SCF) along with MLP classifier for their
work and achieved maximum of 93% recognition accuracy
for 50-city names (7500 words as dataset) written in Bangla
script. Malakar et al. [17] developed a hierarchical feature
selection (HFS) model based on genetic algorithm so that
extracted local and global features of handwritten word
images can be optimized. Authors used database consisting
of 12,000 samples (Bangla words) for developing feature
descriptors depending upon the shape/texture of the hand-
written text images. They reduced the feature dimensions up
to 28% and also, improved recognition performance of
HWR by 1.28%. Authors claimed 95.30% recognition
accuracy using gradient-based and elliptical features and
MLP classification method. For identifying and classifying
the Takri Script, Magotra et al. [16] explored a variety of
text segmentation algorithms based on the structural char-
acteristics of the script. They employed various segmenta-
tion techniques of Gurmukhi characters including touching
and connected component segmentation approaches to test
their suitability for Takri script. Using a Naive-Bayesian
classification technique, authors were able to classify and
identify Takri script texts with an accuracy of 80%.
Kaur and Kumar [11] have developed a holistic approach
based on eXtreme Gradient Boosting (XGBoost) for offline-
HWR (Gurumukhi words). They extracted various features,
namely zoning, diagonal, intersection and open-end points,
and peak extent features from a database of 40,000 samples
of Gurumukhi words and classified them based on XGBoost
approach. They obtained 91.66% accuracy for their pro-
posed system. Narang et al. [21] have proposed a system for
the recognition of ancient texts written in Devanagari script
using a deep learning model (CNN). For the experimental
study, authors have considered a dataset with 5484 char-
acters. Authors achieved the recognition accuracy for
ancient Devanagari characters of 93.73% using CNN as a
feature extractor. In order to effectively recognize Devana-
gari handwritten writing, Pande et al. [24] in their work
configured a Convolutional Neural Network (CNN) using
the most appropriate strategies to increase recognition rate.
Authors achieved 99.13% of accuracy for 46 number of
classes (Devanagari characters) using 12 Layer CNN
approach including Dropout Layer for the recognition of
Devanagari handwritten images (text). Malakar et al. [19]
developed a model that integrates shape transformation
characteristics with a majority voting mechanism to recog-
nize handwritten words (both Arabic and Roman). In order
to get better recognition results, authors additionally took
123
7512
into account inter-segment similarity. In comparison with
cutting-edge holistic word recognition techniques, they
achieved encouraging recognition results for their work.
Deore [3] had proposed a technique for the recognition of
handwritten Devanagari words based on scan profile and
sliding window approaches. Author explored ResNet as
classifier and achieved 86% of accuracy for his work. Singh
et al. [38] presented a board analysis of various feature
extraction and classification techniques developed so for the
recognition of both online as well as offline handwritten
characters/words. Authors discussed most recent studies,
research gaps, difficulties, and prospects in the field of
Devanagari character/word recognition. They summarized
that, nowadays, researchers are focusing on deep learning
techniques in order to attain improved recognition results.
2.1 Previous research outcomes
It has been gathered from the literature that from past few
years, holistic or segmentation-free based approaches are
achieving much attention of researchers working in the area
of handwritten word recognition as compared with analyt-
ical or segmentation based approaches. This is due to better
system performance of holistic approaches over analytical
approaches. Some of the previous research outcomes based
on the state-of-the-art analysis are given below:
● In word recognition of Indic scripts as mentioned in the
above section, most of the authors used HMM, SVM and
MLP classification methods.
● There is a lack of standard database for Devanagari
handwritten words so it arises the difficulty while
analyzing the performance comparison of methods/
systems developed by different researchers working in
this area.
● To get improved recognition accuracy combination of
various features and classification algorithms play an
important role.
● Performance comparison of proposed systems can also
be done in terms of other metrics like F1-score, FAR
(false acceptance rate), FRR (false rejection rate), MCC
(Matthews correlation coefficient) and AUC (area under
the curve) along with recognition accuracy.
● Ensemble methods (like XGBoost, AdaBoost and gra-
dient boosting) can also be used to improve recognition
results by constructing multiple-models and then com-
bining them in an appropriate manner. Ensemble meth-
ods may produce more accurate solutions as compared
with single model.
123
Neural Computing and Applications (2023) 35:7509–7527
3 General framework for handwritten word
recognition
The experiment is performed using various feature extrac-
tion and classification algorithms for a database of hand-
written Devanagari words having 50-word classes (400
sample-images per class or word). A comparative study has
been carried out on the basis of various performance met-
rics, namely recognition accuracy, F1-score, false accep-
tance rate (FAR), false rejection rate (FRR), Matthews
correction coefficient (MCC) and area under the curve
(AUC). Further, to improve recognition accuracy of the
system combination of various feature extraction techniques
has been studied in this work. The present study also reveals
that a combination of uniform zoning-, diagonal- and cen-
troid-based features with gradient boosted decision tree
classifier is showing an acceptable recognition accuracy as
compared to other mentioned techniques. Handwritten word
recognition system for Devanagari script is implemented as
depicted below (Fig. 1):
● Collecting the scanned handwritten word images as
database (or samples).
● Splitting the database into training and test samples.
● Applying preprocessing techniques such as binarization,
normalization and thinning.
● Generating feature vectors by extracting features from
training and test samples using uniform zoning-, diag-
onal- and centroid-based methods along with their
different combinations.
● Classifying the training samples into 50-classes using
K-nearest neighbor, decision tree, random forest and
gradient tree boosted decision tree classification
methods.
● Evaluating the performance of proposed handwritten
word recognition system by predicting the labels/classes
of input test sample images using various performance
metrics.
3.1 Database collection
Database collection is the most tedious phase in offline
handwritten word recognition system as there is a lack of
standard dataset. We have collected handwritten samples of
20,000 Devanagari words as mentioned by Shaw et al. [30–
34] written by hundred of writers in offline mode. This
dataset has been collected from various places including
government & private schools & colleges, government
offices and some public places so that different writing
styles would be captured. This database consists of 50
Neural Computing and Applications (2023) 35:7509–7527
Fig. 1 Block diagram representation of handwritten word recognition system
city/place names (50-word classes) written by different
writers having different geographical background, occupa-
tion, age and academic-qualifications. Some samples of
collected handwritten Devanagari words are shown in
Fig. 2.
This dataset has been partitioned in the ratio of 70% data
(14,000 words) as training and 30% data (6,000 words) as
test samples for experimentation.
3.2 Digitization and preprocessing
Digitization is carried out to transforming the paper-based
documents into digital form by scanning these documents at
300dpi (dots per inch) resolution. After scanning, words as
digital images were stored in a.jpeg image format. There-
after, pre-processing operations such as binarization using
(Otsu [23]), normalization and thinning using [15] were
carried out. Binarization converts the scanned word images
into black and while pixels so as to reduce computational
complexity, whereas, normalization is used to achieve uni-
formity in handwritten word sizes. In this work, a uniform
size of 256
64 has been considered due to horizontal
writing style of Devanagari script. To minimize the text
width from multiple pixels to unit-pixel, thining opertaion is
applied on normalized words which helps to reduce the
amount of data to represent or store a word [41].
3.3 Feature extraction techniques
After preprocessing, feature extraction is an important step
of the handwritten word recognition system. It is used to
Fig. 2 Some samples of collected handwritten words
7513
measure the relevant shape of the character/word so that
recognition rate can be maximized. Generally, the perfor-
mance of the recognition system highly depends on the
quality of features which were being extracted [3]. It
becomes a difficult and challenging task as the nature of
handwriting may include high degree of variability and
roughness. So, in many application including optical char-
acter/word recognition, it becomes essential to extract only
those features that shall be used to differentiate between all
the character/word classes that exist in a particular recog-
nition system. Three statistical feature extraction tech-
niques, namely uniform zoning features, diagonal features
and centroid features, have been considered for this work.
Also, performance analysis of various combinations of
these techniques has been carried out in this presented work.
3.3.1 Uniform zoning features
In uniform zoning feature extraction method, a handwritten
scanned image is divided into equal sized uniform zones
ðn ¼ 64Þ, as depicted in Fig. 3. Thereafter, the number of
foreground pixels ðp1 ; p2 ; . . .. . .::; pn Þ in each zone are cal-
culated for all n zones and normalized to ½0; 1 so that
resultant feature set of n elements can be obtained [12].
In the present work, a scanned handwritten word image
of sized 256
64 (refer to Fig. 3a) is divided into 4 equal
zones (refer to Fig. 3b) and further, each zone is divided into
16 zones (refer to Fig. 3c). Thereafter, each block among
these 16 zones is further divided into 4 zones (refer to
Fig. 3d). So, total 85 (1?4?16?64) zones of a hand-
written word image are used to calculate its foreground
pixel density.
Steps involved to extract uniform zoning features are as
follows:
Step 1 Give digitized image of a word as input and
convert it into the binary image.
Step 2 Divide the bitmapped word image into n number
of equal sized zones in hierarchical order.
Step 3 Count black pixels (foreground pixels) in each
zone and calculate density of the zone.
123
7514 Neural Computing and Applications (2023) 35:7509–7527

Fig. 3 Various uniform zones of

a word “आसनसोल” (Asansol)
written in Devanagari script: a A
single uniform zone considering
the whole word image; b 1×4
zones of a word image; c 2×8
zones of a word image; and
d 4×16 zones of a word image

D Step 2 After preprocessing, divide it into n number of

Density of the zone ¼ zones (here n ¼ 64 with a pixel size of each zone
2
16
16) in hierarchical order.
where D represents total number of foreground pixels
Step 3 Each zone has m number of diagonals (here,
(black pixels) in a zone) and 2 represents total number of
m ¼ 31). Calculate number of foreground pixels present
pixels in a zone.
along with each diagonal, so that sub-feature values of
Step 4 Store the density value of each zone into a feature
each zone can be extracted.
vector.
Step 4 These m sub-features values are averaged to form a
Finally, 85 features of a handwritten word image are single value so that same can be placed in the
extracted. corresponding zone as its feature. If there is no
foreground pixel in the diagonal, then the value of sub-
3.3.2 Diagonal features feature of that particular zone is considered as zero.
Step 5 Finally, calculate average featured value by
Diagonal-based features play a significant role for achieving summing up all the above obtained sub-feature values
higher recognition accuracy [12, 26]. In this method, firstly of each diagonal.
each scanned handwritten word image is divided into Step 6 Sum up all the sub-feature values of each diagonal
number of equal sized zones. Thereafter, features are and calculate average feature values of all sub-values to
extracted by moving along its diagonal as depicted in Fig. 4. form a single feature value for the particular zone.
Here, feature vector is calculated on the basis on pixels
For this work, 85 features of a handwritten word image
around each diagonal.
are extracted by considering above steps.
The following are the steps used for extracting the above-
mentioned features:
3.3.3 Centroid-based features
Step 1 Input the scanned image of handwritten Devana-
gari word and convert it into corresponding binary image.
In centroid-based feature extraction method, preprocessed
scanned image of handwritten word is divided into n

Fig. 4 Diagonal-based features

of a Devanagari handwritten
word

123
Neural Computing and Applications (2023) 35:7509–7527
number of zones in similar manner as the case of zoning
method. After it, centroid of foreground pixels of each zone
are calculated [13]. Thereafter, zone-wise distances of each
pixel from the centroid are computed. After that, sub-feature
value is obtained by average of all the calculated distances.
Coordinates of the each centroid may also be considered as
one of the vectors in feature set. If a zone has not any
foreground pixel, then zero is assigned as feature vector
value of corresponding zone. Centroid-based features of a
Devanagari handwritten word are depicted in Fig. 5.
Steps for the same has been outlined below:
Step 1 Input the digitized image of a handwritten word
and obtain its binary image.
Step 2 Then, divide the bitmapped word image into n
number of equal sized zones.
Step 3 Calculate the centroid of each zone.
Step 4 Obtain the mean distance of each pixels from the
centroid in a zone and save it to form a sub-feature for
that zone.
Step 6 Obtain a feature vector by repeating above steps
for every zone of an image.
Step 7 Feature value can be taken as zero for those zones
which does not contain any black pixels.
For this work, 85 features as feature vector of a word
image are extracted by applying above mentioned steps.
3.4 Classification techniques
After getting feature vectors through feature extraction step,
further classification acts as a decision making step so as to
recognize the class of unknown handwritten word based on
those extracted features. For any classifier to identify the
unknown handwritten word, initially, the classifier has to be
trained with a certain set of training dataset so that the
classifier must be able to match the extracted features with
preset rules. In this work, three classifiers, namely KNN (K-
nearest neighbor), decision tree and random forest, are used.
These classifiers are briefly outlined in the below sub-
sections.
Fig. 5 Centroid-based features of a Devanagari handwritten word
7515
3.4.1 K-nearest neighbor (KNN)
K-nearest neighbor (KNN) is a supervised, nonparametric,
multi-functional machine learning algorithm that can be
used in various application areas including handwritten
character/word recognition (Abu et al., 2019; [40]. As its
name indicates, it recognize the class of new input data
point by considering ‘K’ nearest neighbors or feature sim-
ilarity or data points. Here, ‘K’ represents the number of
sub-feature vector that has to be stored in a feature vector as
a deciding factor for recognition of a pattern/shape of
unknown handwritten word. For this work, value of ‘K’ is
taken as 1, i.e., single nearest neighbor. The algorithm
computes the neighborhood or distance of an unknown
input data from the target function based on the local
minimum. Distance can be calculated using various distance
metric such as Euclidean distance, Manhattan distance,
Minkowski distance, Hamming distance. In this work, the
Euclidean distance ‘d 0 has been measured using the fol-
lowing equation:
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
XN  2
d¼ j¼1
x j  y j
where j represents the number of features containing in the
feature set and x and y are the features stored in training data
and candidate feature vector. Nearest feature vector having
minimum Euclidean distance shall vote to the candidate
feature for recognizing its class. It has advantage of
robustness where the training data are noisy. Figure 6
depicts the case of KNN classification for K=1, where a red
square (Class-1) is the nearest neighbor to the test sample
that falls inside the dotted circle.
Consider, D is the dataset and an integer K represents the
minimum number of near neighbors that can be taken to
establish proximity. The followings are the steps to predict
the output y for a new observation x:
Step 1 Load the training data.
Fig. 6 KNN classification (K=1)
123
7516
Step 2 Organize the data by applying suitable required
techniques such as scaling, missing-value treatment and
dimensionality reduction.
Step 3 Find the optimal value for K. Here, it is considered
as K=1.
Step 4 Select the K number of the neighbors.
Step 5 Predict a class value for new data:
(a) Calculate the Euclidean distances between the x ob-
servable and all the data points of K number of
neighbors.
(b) Take the K observations that constitute the smaller
distances to the observable point x.
(c) Among these K neighbors, count the number of the
data points in each category.
(d) With the y outputs taken from the K observations,
use the mode of y deductions.
(e) The final prediction shall be done based on the
most frequent class as calculated in above steps.
3.4.2 Decision tree (DT)
The decision tree is another supervised learning-based
machine learning algorithm that is generally used to rec-
ognize the class of a pattern and find its application in
classification problems like handwritten character/word
recognition. As its name decision tree indicates, the algo-
rithm is based on breaking down the data by asking a series
of questions to make a particular decision. It has tree like
graph, where a question is asked at each node. Based on the
feature set received in input, the decision tree tries to match
the pattern accordingly. This classifier continue its classifi-
cation on a particular problem, in sub-parts until it reaches
to the final conclusion. Among the n traits, one trait is
chosen at root node and rest traits are set at diverse levels of
tree nodes. Determination of root trait could be a tiresome
assignment, and it may deliver poor recognition outcomes if
root node is taken as random. Refer to Fig. 7 as an example,
where a decision tree can be used to recognize the particular
class of handwritten Devanagari word (HDW) such as class-
1, class-2 and so on.
For this work, C4.5 decision tree algorithm has been
implemented developed by Quinlan as it is efficient and
powerful [2]. Moreover, it has capability to deal continuous
and categorical values. Mainly, decision tree C4.5 algorithm
consists: choice of trait as a root, generates branch, puts
dataset in branch and repeat the process until each class
shall have the same value. The trait with the highest
information gain value can be used as the root trait and
equation for the same is as follows:
123
Neural Computing and Applications (2023) 35:7509–7527
 
m  
X Cj
Information GainðC; T Þ ¼ Entropy ðC Þ 
jC j
  j¼1

Entropy Cj
 
where C represents set of case; T is a trait of the case; C j  is
the number of cases to j; jC j is the number of cases in the set
and EntropyðCÞ can be calculated as given below:
X
m
Entropy ðC Þ ¼  pj
logpj
j¼1
where p represents the fraction of traits in a particular class.
The pseudo-code for decision tree C4.5 algorithm is as
follows:
3.4.3 Random forest (RF)
Random forest is another most commonly used ensemble
type supervised machined learning algorithm due to its
accuracy, simplicity and flexibility [13]. It combines mul-
tiple algorithms to produce improved classification results
rather than using any individual algorithm (refer to Fig. 8).
It is capable of constructing multiple decision trees
during training time. Detection of root node and splitting of
feature node occurs randomly, in this type of classification.
It helps to avoid the problem of over-fitting as may occur in
decision tree. Random forest produces as class of pattern
which is the resultant average of all possible outputs that
shall be received from the multiple decision trees. RF
classifier randomly chooses features and gives interpreta-
tions accordingly while decision tree tends to produce/cre-
ate rules for the same. Figure 8 depicts the process of
classification using random forest. Due to its efficient per-
formance in the area of pattern recognition, this classifier is
Neural Computing and Applications (2023) 35:7509–7527
Fig. 7 Decision Tree (DT)
Classification
Fig. 8 Random forest classification
being used in this work to analyze the performance of
handwritten Devanagari word recognition.
3.4.4 Gradient boosted decision tree (GBDT)
Gradient boosted decision tree (GBDT) is a machine
learning method that can be used for combining the weak
classifiers for achieving a strong ensemble [22, 41]. The
weak learners may be the individual decision trees itself.
This method optimizes the predictive value of a model by
sequential steps during training or learning. In decision
7517
tree, each iteration adjusts the values of its weights that
are applied to each input variables for predicting the
target value so that loss function can be minimized. In
this, all trees are connected in sequential manner and tries
to minimize the error of the previous tree as depicted in
Fig. 9. Thus, this method improves the training or
learning process through simplifying the objective so to
obtain suitably optimal solution. Finally, the model
aggregates the prediction or result of each stage and thus
a strong learner for handwritten word recognition can be
achieved.
123
7518 Neural Computing and Applications (2023) 35:7509–7527

Fig. 9 Work flow of Gradient Boosted Decision Tree (GBDT)

4 Experiment results and discussion Number of correct predictions

RA ¼
100
Number of input data samples
In the section, experiment results based on some statistical
features and classification methods considered have been
4.1.2 F1-score
presented in terms of various performance metrics which
are briefly defined in the following subsection.
It indicates the harmonic average of the precision and recall.
Its best value is 1 and worst is 0. For multi-class problems,
4.1 Performance metrics
it is calculated as weighted average of F1-score of each
class.
The performance of the proposed recognition system has  
been evaluated and analyzed in terms of various perfor- precision
recall
F1-score ¼ 2

mance metrics, namely recognition accuracy, F1-score, false precision þ recall
acceptance rate, false rejection rate, Matthews correction
Also,F1  score ¼ TPþ1ðTP
FPþFN Þ
.
coefficient and area under the curve. These metrics can be 2

evaluated using confusion matrix [11, 29]. Important terms

4.1.3 False acceptance rate (FAR)
related to the confusion metrics are listed below:
True Positive (TP): It represents correct number of pos-
False acceptance rate (FAR) can be defined as the per-
itive predictions.
centage of identification instances in which unauthorized
True Negative (TN): It gives correct number of negative
instances are wrongly or incorrectly accepted. It can be
predictions.
expressed as the division of the number of false acceptances
False Positive (FP): It specifies incorrect number of
and the number of identification attempts. Mathematically,
positive predictions.
FAR can be given as:
False Negative (FN): It indicates incorrect number of
negative predictions. FP
FAR ¼
100
Various performance metrics used for this work are ðFP þ TNÞ
outlined below.
4.1.4 False rejection rate (FRR)
4.1.1 Recognition accuracy
False acceptance rate (FAR) can be defined as the per-
Recognition accuracy (RA) can be defined as the ratio of the
centage of identification instances in which unauthorized
number of correctly recognized test samples to the total
instances are wrongly rejected. It can be expressed as the
number of input or test data samples [5].
division of the number of false rejections and the number of

123
Neural Computing and Applications (2023) 35:7509–7527 7519

identification attempts. Mathematically, FRR can be repre- distinguishing between positive and negative classes. For
sented as: binary classification, it can be expressed as follows [39]:
FN 
FRR ¼
100 1 TP TN
ðFN þ TPÞ AUC ¼ þ
2 TP þ FN TN þ FP
The system performance of proposed work has been
4.1.5 Matthews correction coefficient (MCC)
discussed in the following subsection based upon above
mentioned performance metrics.
Matthews correlation coefficient (MCC) uses entire entries
of confusion matrix. It is considered as better singular
performance metric specially suited for imbalanced data 4.2 Experimental results
training. Mathematically, it can be expressed as [9]:
ðTP
TNÞ  ðFP
FNÞ Performance metrics-wise handwritten word recognition
MCC ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi results and analysis of the proposed system have been
ðTP þ FPÞðTP þ FNÞðTN þ FPÞðTN þ FNÞ
presented in the various tables (refer to Tables 1, 2, 3, 4, 5,
6) for features extraction and classification techniques used
4.1.6 Area under the curve (AUC) for this work. These results are also represented graphically
for better understanding (see Figs. 10, 11, 12, 13, 14 and
Area under the curve (AUC) represents the ability of a 15).
classifier to avoid wrong classification. Higher value of
AUC indicates the better performance of the system for

Table 1 Recognition accuracy (%) results with 70%-30% partition strategy

Feature extraction technique Classifier-wise recognition accuracy (in terms of %) (70% data as training set and 30% data as testing set)
K-nearest neighbors Decision tree Random forest Gradient boosted decision tree
(KNN) (DT) (RF) (GBDT)

Uniform Zoning Features 70.86 75.33 79.68 82.40

(FE1)
Diagonal Features (FE2) 71.91 76.50 81.01 83.25
Centroid Features (FE3) 73.85 78.00 81.41 84.03
FE1?FE2 84.50 87.41 89.60 92.16
FE1?FE3 85.41 87.90 90.03 92.98
FE2?FE3 86.10 88.58 90.85 93.95
FE1?FE2?FE3 86.38 89.18 91.60 94.53

Table 2 F1-score results with 70%-30% partition strategy

Feature extraction technique Classifier-wise F1-score (in terms of %) (70% data as training set and 30% data as testing set)
K-nearest neighbors Decision tree Random forest Gradient boosted decision tree
(KNN) (DT) (RF) (GBDT)

Uniform Zoning Features 71.19 75.56 79.76 83.32

(FE1)
Diagonal Features (FE2) 72.20 76.68 81.02 83.21
Centroid Features (FE3) 74.08 78.14 81.39 84.00
FE1?FE2 84.46 87.43 89.63 92.22
FE1?FE3 85.42 87.93 90.05 93.05
FE2?FE3 86.08 88.61 90.88 94.00
FE1?FE2?FE3 86.36 89.22 91.66 94.56

123
7520 Neural Computing and Applications (2023) 35:7509–7527

Table 3 False acceptance rate (FAR) results with 70%-30% partition strategy
Feature extraction technique Classifier-wise false acceptance rate (FAR) (in terms of %) (70% data as training set and 30% data as testing set)
K-nearest neighbors Decision tree Random forest Gradient boosted decision tree
(KNN) (DT) (RF) (GBDT)

Uniform Zoning Features 0.59 0.50 0.41 0.35

(FE1)
Diagonal Features (FE2) 0.57 0.47 0.38 0.34
Centroid Features (FE3) 0.53 0.44 0.37 0.32
FE1?FE2 0.31 0.25 0.21 0.15
FE1?FE3 0.29 0.24 0.20 0.14
FE2?FE3 0.28 0.23 0.18 0.12
FE1?FE2?FE3 0.27 0.22 0.17 0.11

Table 4 False rejection rate (FRR) results with 70%-30% partition strategy
Feature extraction technique Classifier-wise false rejection rate (FRR) (in terms of %) (70% data as Training set and 30% data as testing set)
K-nearest neighbors Decision tree (DT) Random forest (RF) Gradient boosted decision tree (GBDT)
(KNN)

Uniform zoning features (FE1) 29.13 24.66 20.31 17.60

Diagonal features (FE2) 28.08 23.50 18.98 16.74
Centroid features (FE3) 26.15 22.00 18.58 15.96
FE1?FE2 15.50 12.58 10.40 7.83
FE1?FE3 14.58 12.10 9.96 7.01
FE2?FE3 13.90 11.41 9.15 6.05
FE1?FE2?FE3 13.61 10.81 8.40 5.46

Table 5 Matthews correction coefficient (MCC) results with 70%-30% partition strategy
Feature extraction technique Classifier-wise Matthews correction coefficient (MCC) (70% data as Training Set and 30% data as Testing Set)
K-nearest neighbors Decision tree (DT) Random forest (RF) Gradient boosted decision tree (GBDT)
(KNN)

Uniform zoning features (FE1) 0.709 0.754 0.798 0.825

Diagonal features (FE2) 0.720 0.766 0.811 0.833
Centroid features (FE3) 0.739 0.781 0.815 0.841
FE1?FE2 0.845 0.875 0.896 0.921
FE1?FE3 0.855 0.879 0.900 0.929
FE2?FE3 0.861 0.886 0.908 0.939
FE1?FE2?FE3 0.864 0.892 0.916 0.945

4.2.1 System performance based on recognition accuracy
Experimental results in terms of recognition accuracy (RA)
for various feature extraction and classification techniques
considered for this experimental work are depicted in
Table 1.
The following figure (Fig. 10) gives graphical represen-
tation for the same.
It has observed that using GBDT classification along
with combination of various feature extraction techniques,
namely uniform zoning-, diagonal- and centroid-based
features (FE1?FE2?FE3), maximum recognition accuracy
(RA) of 94.53% has been obtained. It can also be seen from
table that there is the improvement in recognition accuracy
when combination of various statistical features have been
considered. Minimum recognition accuracy of 70.86% is
obtained by considering zoning feature extraction and KNN
classification techniques. Similar results have been obtained
by Kaur and Kumar, (11, 12 in their work for handwritten
recognition of Gurmukhi words.

123
Neural Computing and Applications (2023) 35:7509–7527 7521

Table 6 Area under the curve (AUC) results with 70%-30% partition strategy
Feature extraction technique Classifier-wise area under the curve (AUC) (in terms of %) (70% data as Training Set and 30% data as testing set)
K-nearest neighbors Decision tree (DT) Random forest (RF) Gradient boosted decision tree (GBDT)
(KNN)

Uniform zoning features (FE1) 85.13 87.41 89.63 91.02

Diagonal features (FE2) 85.67 88.01 90.31 91.45
Centroid features (FE3) 86.65 88.77 90.51 91.85
FE1?FE2 92.09 93.57 94.69 96.00
FE1?FE3 92.55 93.82 94.91 96.42
FE2?FE3 92.90 94.17 95.33 96.91
FE1?FE2?FE3 93.05 94.48 95.71 97.21

Fig. 10 Recognition accuracy based on various feature extraction and classification methods considered

Fig. 11 F1-score based on various feature extraction and classification methods considered

123
7522 Neural Computing and Applications (2023) 35:7509–7527

Fig. 12 False acceptance rate (FAR) based on various feature extraction and classification methods considered

Fig. 13 False rejection rate (FRR) based on various feature extraction and classification methods considered

4.2.2 System performance based on F1-score 4.2.3 System performance based on FAR

In Table 2, comparative analysis of various feature extrac-
tion and classification techniques explored for HWR of
Devanagari script is represented in terms of F1-score. These
results are also presented graphically in Fig. 11, which
clearly indicates the performance of the proposed frame-
work. Authors have observed that (FE1?FE2?FE3) fea-
tures and GBDT classification is the better combination as
maximum F1-score of 94.56% is obtained. Because F1-
score is the mean of precision and recall, it gives the same
amount of weight for both precision and recall. High F1-
score indicates that the system has high precision and recall
values.
The maximum FAR achieved is 0.59% based on uniform
zoning features (FE1) and KNN classification, whereas
minimum FAR obtained is 0.11% from combination of all
three features and GBDT classification techniques as
depicted in Table 3. Low value of FAR is desirable for
handwritten word recognition applications. Low values of
FAR give rise to higher value of FRR. False acceptance rate
(FAR) results are also presented graphically in Fig. 11 for
the framework taken for this work.

123
Neural Computing and Applications (2023) 35:7509–7527 7523

Fig. 14 Matthews correction coefficient (MCC) based on various feature extraction and classification methods considered

Fig. 15 Area under the curve (AUC) based on various feature extraction and classification methods considered

4.2.4 System performance based on FRR 4.2.5 System performance based on MCC

GBDT classifier scored minimum FRR of 5.46% by con-
sidering the combination of all features (FE1?FE2?FE3)
for HWR system developed for Devanagari script as given
in Table 4. FAR and FRR values depend upon each other.
As FAR decreases, FRR increases and vice versa. If the
values of both metrics (FAR and FRR) are same, then there
exists a point where lines intersect to each other called equal
error rate (ERR). False rejection rate (FRR) results are also
presented graphically in Fig. 12 for the framework taken for
this work (Figs. 13, 14).
Maximum achieved MCC value is 0.945 (see Table 5) for
combination of three features, namely uniform zoning-,
diagonal-, and centroid-based features using GBDT classi-
fier, whereas minimum attained MCC value is 0.709 by
considering uniform features only along with KMM
classifier.
4.2.6 System performance based on AUC
Similarly, maximum achieved AUC value is 97.21% (refer
to Table 6) for combination of three features, namely uni-
form zoning-, diagonal- and centroid-based features using

123
7524 Neural Computing and Applications (2023) 35:7509–7527

GBDT classifier. Higher values of AUC indicate the better
performance of the system.
In Fig. 15, graphical representation of area under the
curve (AUC) is depicted based on various feature extraction
and classification methods considered for handwritten
Devanagari word recognition.
Experimental study shows that by considering combi-
nation of three features, namely uniform zoning-, diagonal-,
and centroid-based features (FE1?FE2?FE3) along with
GBDT classifier improve the recognition accuracy and other
performance metrics for handwritten Devanagari word
recognition system.
4.3 Comparison with existing methods
and syntactic analysis
In this section, syntactic analysis as a comparison among
exiting state-of-the-art work and proposed work is carried
out. The proposed system has achieved recognition accu-
racy of 94.53%, F1-score of 94.56%, FAR of 0.11%, FRR
of 5.46%, MCC of 0.945 and AUC of 97.21%. It can be
observed from Table 7 that our proposed system obtained
better and competent performance as compared with other
similar systems. It is also to mention that due to the lack of
standard database, researchers had performed their experi-
mentation on collected corpus of handwritten words. Hence,
it is very challenging and difficult to directly compare the
recognition results. Parui and Shaw [25] achieved recogni-
tion accuracy of 87.71% on a corpus of 10,000 handwritten
words of Devanagari script. In [33, 34], authors obtained
recognition accuracies of 80.2% and 84.31% using direc-
tional chain code and stroke based features, respectively, by
considering HMM classifier for each and a corpus of 39,700
handwritten words (Devanagari). Using a corpus of 13,000
handwritten Devanagari words, Shaw and Parui [30]
attained 91.25% of recognition accuracy. They explored
stroke and wavelet-based features with HMM classification.
Singh et al. [35] in their work achieved maximum
recognition accuracy of 93.21% based on curvelet transform
feature set and KNN classification on a corpus of 28,500

Table 7 Comparison of proposed work with existing methodologies

Researchers Recognized indic Dataset Methodology Recognition
scripts (Words) accuracy (%)
Feature set Classifier

Parui and Shaw Devanagari 10,000 Stroke based HMM 87.71

[25]
Shaw et al. [33] Devanagari 39,700 Directional chain code HMM 80.2
Shaw et al. [34] Devanagari 39,700 Stroke based HMM 84.31
Shaw and Parui, Devanagari 13,000 Stroke based (Stage-1); Wavelet (Stage-2) HMM (Stage-1); Modified 91.25
[30] Byes (Stage-2) (Stage-2)
Singh et al. [35] Devanagari 28,500 Curvelet transform SVM and KNN 85.6 (SVM);
93.21
(KNN)
Ramachandrula Hindi 39,600 Directional element Dynamic Programming 91.23
[27]
Shaw et al. [31] Devanagari 39,700 Combination of skeleton and contour based SVM 79.01
Shaw et al. [32] Devanagari 39,700 DDD and GSC Multi-class SVM 88.75
Kumar [14] Devanagari More than Neighbor pixel weight and gradient feature MLP 80.8
3500
Malakar et al. Hindi 4620 Low-level features MLP 96.82
[20]
Bhunia et al. [1] Bangla, 3856; 3589 PHOG feature HMM (for Middle-zone) Above 60
Devanagari and and 3142 SVM (for Upper/Lower
Gurumukhi zone)
Ghosh et al. [5] Bangla 7500 Gradient features and modified SCF; MA- MLP 93
based wrapper filter selection approach
Malakar et al. Bangla 12,000 Gradient-based and elliptical MLP 95.30
[17, 18]
Kaur and Gurumukhi 40,000 Zoning features XGBoost 91.66
Kumar [11]
Proposed Study Devanagari 20,000 Combination of uniform zoning, diagonal Gradient boosted decision 94.53
and centroid features tree

123
Neural Computing and Applications (2023) 35:7509–7527
Fig. 16 Confusion matrix
handwritten words (Devanagari). Ramachandrula et al. [27]
obtained maximum recognition accuracy of 91.23% by
considering directional element based features and dynamic
programming on a database of 39,600 Hindi words. Fur-
thermore, in [31, 32], authors obtained recognition accura-
cies of 79.01% (skeleton and contour based feature set) and
88.75% (DDD and GSC feature set) by considering SVM
and multi-class SVM classification methods 39,700 hand-
written words (Devanagari). Kumar [14] got recognition
accuracy of 80.8% (neighbor pixel weight and gradient
feature set) on a corpus of more than 3500 Devanagari
words using MLP classifier. Further, based on MLP clas-
sification, Malakar et al. [20] explored low-level features
7525
and obtained recognition accuracy of 96.82% (corpus of
4620 Hindi words). Bhunia et al. [1] achieved recognition
accuracy of above 60% using PHOG feature set and HMM
& SVM classification on a database of 3856 (Bangla), 3589
(Devanagari) and 3142 (Gurumukhi) words. On corpus of
7500 handwritten Bangla words, Ghosh et al. [5] obtained
93% of recognition accuracy by exploring gradient and
modified SCF features and MA-based wrapper filter selec-
tion approach along with MLP classification, whereas on a
database of 12,000 handwritten Bangla words, Malakar
et al. [17, 18] obtained higher recognition accuracy of
95.30% based on gradient-based and elliptical feature set
and MLP classifier. Moreover, Kaur and Kumar, [11]
123
7526
obtained recognition accuracy of 91.66% (on a database of
40,000 handwritten Gurumukhi words) by exploring zoning
features and XGBoost. It is evident from experimentation
and syntactic analysis that authors achieved better recog-
nition results by incorporating gradient boosted decision
tree with the combination of statistical features for recog-
nition of handwritten word recognition system for
Devanagari words. In Fig. 16, the confusion matrix is
depicted when a combination of zoning-, diagonal- and
centroid-based features along with GBDT classifier has
been considered.
It shows correlation amongst predicted class (word) and
actual class (word). It is a useful tool for analyzing the
effectiveness of a classification phase of handwritten word
recognition system. A confusion matrix is an N
N matrix,
where N denotes target classes, i.e., predicted and actual
classes (for this work, N ¼ 50). It is usually used to visu-
alize and summarize the performance of a classification
algorithm, i.e., how well a classification algorithm performs.
Every entry in a confusion matrix indicates whether the
proposed system recognize the word-classes correctly or
incorrectly. As a result, the confusion matrix makes it evi-
dent how the system is doing while categorizing or recog-
nizing various word-classes. Thus, confusion matrix is
significant tool as it provides comprehensive understanding
(it compares the actual word-classes with predicted word-
classes) of how well the proposed system is performing.
The confusion matrix of size 50
50(refer to Fig. 16),
summarizes the experimental results for handwritten
Devanagari word recognition system (for each of 50 word-
classes considered for this work). Each row represents the
frequency count of handwritten words recognized or iden-
tified by the system with darker colors associated with
higher frequency counts. The handwritten Devanagari
words can either be recognized correctly (predicted word-
class is same as actual word-class) or incorrectly (predicted
word-class is not same as actual word-class). In this work,
maximum recognition accuracy of 94.53% has been com-
puted from the confusion matrix.
5 Conclusions and future prospects
Recognition of Indic and non-Indic handwritten words has
been an active and popular research area from past few
years due to its variety of real-time potential applications
including postal automation, etc. The aim of this article is to
present a performance-based analytical study of different
features and classifiers for the recognition of handwritten
Devanagari words. In this article, unıform zoning-, diago-
nal- and centroid-based features are explored along with
various classifiers, namely KNN, decision tree, random
forest and gradient boosted decision tree for experimental
123
Neural Computing and Applications (2023) 35:7509–7527
work. Experimentation is also performed by considering
combination of the above-mentioned features and classi-
fiers. Based on the experimental study, maximum recogni-
tion accuracy of 94.53%, F1-score of 94.56%, FAR of
0.11%, FRR of 5.46%, MCC of 0.945 and AUC of 97.21%
are achieved for the handwritten Devanagari word recog-
nition. Moreover, comparison with existing methods and
syntactic analysis is presented for the assessment of
recognition performance in terms of the recognition accu-
racy. Overall, proposed system performed good and com-
petent as compared with other existing similar state-of-the-
art systems for recognizing the handwritten Devanagari
words. In the future work, the size of corpus can be
increased for experimental study and to test the effective-
ness of proposed system. The work can be explored for
other similar Indic scripts such as Bangla and Gurumukhi.
Authors contribution Sukhjinder Singh contributed to the database
collection, methodology and original draft. Naresh Kumar Garg
assisted in the supervision, quality check, review and editing. Munish
Kumar Jindal performed the co-supervision, further review and final
editing.
Data availability The datasets generated during and analyzed during
the current study are available from the corresponding author on rea-
sonable request.
Declarations
Conflict of interest The authors declare that they have no conflict of
interest in this work.
References
1. Bhunia AK, Roy PP, Mohta A, Pal U (2018) Cross-language
framework for word recognition and spotting of Indic scripts.
Pattern Recogn 79:12–31
2. Budiman E, Kridalaksana AH, Wati M (2017) Performance of
decision tree C4.5 algorithm in student academic evaluation. In:
Alfred R, Iida H, Ag Ibrahim A, Lim Y (eds) Computational
science and technology (ICCST). Lecture Notes in Electrical
Engineering, vol 488, pp 380–389
3. Deore SP (2022) HDWR_SmartNet: a smart handwritten
Devanagari word recognition system using deep ResNet-based on
scan profile method. Data Sci, pp 57–77
4. Dutta K, Krishnan P, Mathew M, Jawahar C (2017) Towards
accurate handwritten word recognition for Hindi and Bangla. In:
Rameshan R, Arora C, Dutta Roy S (eds) Computer Vision, Pat-
tern Recognition, Image Processing, and Graphics. NCVPRIPG
2017. Communications in computer and information science, vol
841, pp 470–480
5. Ghosh M, Malakar S, Bhowmik S, Sarkar R, Nasipuri M (2019)
Feature selection for handwritten word recognition using memetic
algorithm. In: Mandal J, Dutta P, Mukhopadhyay S (eds)
Advances in intelligent computing. studies in computational
intelligence, vol 687, pp 103–124
Neural Computing and Applications (2023) 35:7509–7527
6. Gupta JD, Samanta S, Chanda B (2018) Ensemble classifier-based
off-line handwritten word recognition system in holistic approach.
IET Image Proc 12(8):1467–1474
7. Gurav Y, Bhagat P, Jadhav R, Sinha S (2020) Devanagari hand-
written character recognition using convolutional neural networks.
In: IEEE International conference on electrical, communication,
and computer engineering (ICECCE), pp 1–6
8. Jayadevan R, Kolhe SR, Patil PM, Pal U (2011) Offline recog-
nition of Devanagari script: a survey. IEEE Trans Syst Man
Cybern Part C (Appl Rev) 41(6):782–796
9. Jurman G, Riccadonna S, Furlanello C (2012) A comparison of
MCC and CEN error measures in multi-class prediction. PLoS
ONE 7(8):1–8
10. Kaur H, Kumar M (2018) A comprehensive survey on word
recognition for non-Indic and Indic scripts. Pattern Anal Appl 21
(4):897–929
11. Kaur H, Kumar M (2021) Offline handwritten Gurumukhi word
recognition using eXtreme gradient boosting methodology. Soft
Comput 25(6):4451–4464
12. Kaur H, Kumar M (2021) On the recognition of offline hand-
written word using holistic approach and AdaBoost methodology.
Multimed Tools Appl 80(7):11155–11175
13. Kaur RP, Kumar M, Jindal MK (2020) Newspaper text recognition
of Gurumukhi script using random forest classifier. Multimed
Tools Appl 79(11):7435–7448
14. Kumar S (2016) A study for handwritten Devanagari word
recognition. In: IEEE International conference on communication
and signal processing (ICCSP), pp 1009–1014
15. Lee TC, Kashyap RL, Chu CN (1994) Building skeleton models
via 3-D medial surface axis thinning algorithms. CVGIP Graph
Models Image Process 56(6):462–478
16. Magotra S, Kaushik B, Kaul A (2020) A comparative analysis for
identification and classification of text segmentation challenges in
Takri Script. Sadhana. https://doi.org/10.1007/s12046-020-01384-4
17. Malakar S, Ghosh M, Bhowmik S, Sarkar R, Nasipuri M (2020) A
GA based hierarchical feature selection approach for handwritten
word recognition. Neural Comput Appl 32(7):2533–2552
18. Malakar S, Paul S, Kundu S, Bhowmik S, Sarkar R, Nasipuri M
(2020) Handwritten word recognition using lottery ticket
hypothesis based pruned CNN model: a new benchmark on
CMATERdb212. Neural Comput Appl 32(18):15209–15220
19. Malakar S, Sahoo S, Chakraborty A, Sarkar R, Nasipuri M (2022)
Handwritten Arabic and Roman word recognition using holistic
approach. Vis Comput. https://doi.org/10.1007/s00371-022-
02500-7
20. Malakar S, Sharma P, Singh PK, Das M, Sarkar R, Nasipuri M
(2017) A holistic approach for handwritten hindi word recogni-
tion. Int J Comput Vis Image Process (IJCVIP) 7(1):59–78
21. Narang SR, Kumar M, Jindal MK (2021) DeepNetDevanagari: a
deep learning model for Devanagari ancient character recognition.
Multimed Tools Appl 80:20671–20686. https://doi.org/10.1007/
s11042-021-10775-6
22. Neelakandan S, Paulraj D (2020) A gradient boosted decision tree-
based sentiment classification of twitter data. Int J Wavelets
Multiresol Inf Process 18(4):205027
23. Otsu N (1979) A threshold selection method from gray-level
histograms. IEEE Trans Syst Man Cybern 9(1):62–66. https://doi.
org/10.1109/TSMC.1979.4310076
24. Pande SD, Jadhav PP, Joshi R, Sawant AD, Muddebihalkar V,
Rathod S, Gurav MN, Das S (2022) Digitization of handwritten
Devanagari text using CNN transfer learning–A better customer
service support. Neurosci Inform 2(3):100016
25. Parui SK, Shaw B (2007) Offline handwritten devanagari word
recognition: an HMM based approach. In: Ghose A, De, RK, Pal
SK (eds) PReMI 2007, LNCS, vol 4815, pp 528–535
7527
26. Pradeep J, Srinivasan E, Himavathi S (2010) Diagonal feature
extraction based handwritten character system using neural net-
work. Int J Comput Appl 8(9):17–22
27. Ramachandrula S, Jain S, Ravishankar H (2012) Offline hand-
written word recognition in Hindi. In: Workshop on document
analysis and recognition, pp 49–54, ACM, New York
28. Roy RK, Mukherjee H, Roy K, Pal U (2022) CNN based recog-
nition of handwritten multilingual city names. Multimedia Tools
Appl 81(8):1–17
29. Sampath A, Gomathi DN (2017) Decision tree and deep learning
based probabilistic model for character recognition. J Central
South Univ 24(12):2862–2876
30. Shaw B, Parui SK (2010) A two stage recognition scheme for
offline handwritten Devanagari words. In: Machine interpretation
of patterns: image analysis and data mining, pp. 145–165, World
Scientific
31. Shaw B, Bhattacharya U, Parui SK (2014) Combination of fea-
tures for efficient recognition of offline handwritten Devanagari
words. In: IEEE International conference on frontiers in hand-
writing recognition, pp 240–245
32. Shaw B, Bhattacharya U, Parui SK (2015) Offline handwritten
Devanagari word recognition: information fusion at feature and
classifier levels. In: IEEE 3rd IAPR Asian conference on pattern
recognition, pp 720–724, IEEE Press, Malaysia
33. Shaw B, Parui SK, Shridhar M (2008) A segmentation based approach
to offline handwritten Devanagari word recognition. In: IEEE Inter-
national conference on information technology, pp. 256–257
34. Shaw B, Parui SK, Shridhar M (2008) Offline handwritten
Devanagari word recognition: A holistic approach based on
directional chain code feature and HMM. In: IEEE International
Conference on Information Technology, pp 203–208
35. Singh B, Mittal A, Ansari M, Ghosh D (2011) Handwritten
Devanagari word recognition: a curvelet transform based
approach. Int J Comput Sci Eng 3(4):1658–1665
36. Singh S, Garg NK (2019) Optical character recognition techniques
for postal automation—a review. In: IEEE International confer-
ence on intelligent computing and control systems (ICICCS),
pp 1385–1389
37. Singh S, Garg NK (2022) Performance analysis of thinning algo-
rithms for offline-handwritten Devanagari words. In: Sengodan T,
Murugappan M, Misra S (eds) Advances in electrical and computer
technologies. ICAECT 2021. Lecture Notes in Electrical Engineer-
ing, vol. 881 https://doi.org/10.1007/978-981-19-1111-8_34
38. Singh S, Garg NK, Kumar M (2022) Feature extraction and
classification techniques for handwritten Devanagari text recog-
nition: a survey. Multimed Tools Appl. https://doi.org/10.1007/
s11042-022-13318-9
39. Sokolova M, Lapalme G (2009) A systematic analysis of perfor-
mance measures for classification tasks. Inf Process Manage 45
(4):427–437
40. Tan S (2006) An effective refinement strategy for KNN text
classifier. Expert Syst Appl 30(2):290–298
41. Wahid MF, Shahriar MF, Sobuj MSI (2021) A classical approach
to handcrafted feature extraction techniques for Bangla hand-
written digit recognition. In: IEEE International conference on
electronics, communications and information technology (ICE-
CIT), pp. 1–4
Publisher's Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Springer Nature or its licensor (e.g. a society or other partner) holds
exclusive rights to this article under a publishing agreement with the
author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such
publishing agreement and applicable law.
123