Online Kanji Characters Based Writer Identification Using Sequential Forward Floating Selection and Support Vector Machine

Abstract

Writer identification has become a hot research topic in the fields of pattern recognition, forensic document analysis, the criminal justice system, etc. The goal of this research is to propose an efficient approach for writer identification based on online handwritten Kanji characters. We collected 47,520 samples from 33 people who wrote 72 online handwritten-based Kanji characters 20 times. We extracted features from the handwriting data and proposed a support vector machine (SVM)-based classifier for writer identification. We also conducted experiments to see how the accuracy changes with feature selection and parameter tuning. Both text-dependent and text-independent writer identification were studied in this work. In the case of text-dependent writer identification, we obtained the accuracy of each Kanji character separately. We then studied the text-independent case by considering some of the top discriminative characters from the text-dependent case. Finally, another text-dependent experiment was performed by taking two, three, and four Kanji characters instead of using only one character. The experimental results illustrated that SVM provided the highest identification accuracy of 99.0% for the text-independent case and 99.6% for text-dependent writer identification. We hope that this study will be helpful for writer identification using online handwritten Kanji characters.

1. Introduction

Writer identification has become a hot research topic in the fields of pattern recognition, forensic document analysis, the criminal justice system, etc. It is one kind of biometric recognition that can easily identify the writer based on their handwritten characters. The handwriting style of each individual, which appears identical at first glance and yet has its own originality, is an interesting research topic of writer identification. Handwritten characters play a significant role for forensic experts in identifying the writers. In the last twenty years, many attempts have been undertaken to use handwritten characters for writer verification and identification. Moreover, handwritten characters can be used for the identification of age groups [1,2]. Writer identifications are utilized for verification and authentication, for example, signature verification and identification in a court of law and bank. With the development of the information society, the importance of writer identification technology to verify the legitimacy of users is increasing. There are three types of authentication methods: knowledge authentication, property authentication, and biometric authentication [3]. The challenges in identifying writers based on handwriting include the following: (i) developing a suitable model to identify the handwriting of different individuals; (ii) extracting and identifying potential handwriting features; (iii) testing the performance of proposed methods. There are numerous studies conducted on writer identification using different language-based characters such as Arabic [4], Bangla [5], Chinese [6,7], French [8], Japanese [8,9,10], and English [11,12,13,14,15].

In the field of handwriting analysis, the goal is to propose an efficient method that can be compared with existing approaches that use veins, fingerprints, and other biometrics. There are two ways of collecting handwritten samples for writer identification: offline and online character recognition methods. In offline-based identification or static method, writing-based samples are collected from documents or images using scanners. Generally, this method is built on several attributes, including lines, characters, words, and so on [16]. The dynamic features are not present in the offline mode. Moreover, there is a lack of sequential information in handwriting and huge intra-class variation. As a result, offline writer recognition is considered more complex. Research on writer identification based on offline methods is a very challenging issue for security and forensic purposes [7]. As a result, many studies have been conducted using offline methods [8,17,18].

Online handwritten characters are considered less challenging due to having various information such as pen pressure, pen altitude, pen azimuth, and pen position [7]. Moreover, handwritten characters or samples can be easily collected using tablets, smart phones, magnetic input, and so on. Writer-identification-based approaches can be categorized into two parts: text-independent and text-dependent. The text-independent methods analyze the entire image without knowing anything about the script’s content, whereas text-dependent approaches analyze the entire image based on such knowledge. The summary of previous research on the identification of writers using handwriting is presented in

Table 1. Previous research on writer identification.

Authors	DT	SS	Methods	Language	Writers	ACC(%)
Nakamura et al. [9]	Online	1230	ANOVA	Kanji	41	90.4
Soma et al. [11]	Offline	5000	Majority Voting	Kanji	100	99.0
Namboodiri and Gupta [13]	Online	400	NN	English	6	88.0
Li et al. [14]	Online	1500	k-NN	English	242	93.6
Wu et al. [15]	Online	1700	HMM	English	200	95.5
Nguyen et al. [17]	Offline	2965	CNN	Kanji	480	93.8
Rehman et al. [6]	Online	52,800	P2P	Chinese	48	99.0
Abdi et al. [4]	Offline	4800	k-NN	Arabic	82	90.2
Nasuno and Arai [10]	offline	50,000	CNN	Japanese	100	90.0

DT—data types; SS—sample size.

. Only a few types of research on writer identification have been conducted on handwritten-based Kanji characters. In some sense, our research focuses on both online text-independent and text-dependent writer identification approaches for handwritten-based Kanji characters. In this study, we use x-coordinates, y-coordinates, pen pressure, pen altitude, pen azimuth, and the time taken to acquire the data. We collected online handwritten-based Kanji characters using a pen tablet. We extracted 47,520 samples from 33 people who wrote different 72 Kanji characters 20 times each and used efficient features to achieve a higher identification rate. Furthermore, we compared how the accuracy varies depending on whether the type is text-dependent or text-independent.

The organization of this paper is as follows: Section 2 presents related work; Section 3 presents the materials and methods, including dataset preparation, proposed methodology, feature extraction, feature normalization, feature selection, and classification using SVM. The experimental results and discussion are discussed in Section 4. Finally, the conclusion and future work direction are discussed in Section 5.

2. Related Work

Writer identification based on online Kanji characters can achieve a greater identification rate than offline Kanji characters because the online characters provide more information such as the length of the stroke order and brushstrokes. We reviewed various existing papers on writer identification, which were mainly conducted on only online data or offline data types based on various handwritten characters, such as English, Arabic, Japanese, Chinese, and so on. We summarized these existing papers and presented in Table 1. Our study mainly focuses on writer identification based on online handwritten Kanji characters. Efficient feature extraction and feature selection play a significant role in writer identification. Various local features were extracted from the characters to show the discriminative information of the writer. For example, Dargan et al. [19] used four features (transition, diagonal, zoning, and peak-extent-based) in order to develop a writer identification system using Devanagari characters. Bensefia and Djeddi [20] proposed a novel feature selection approach for writer identification and obtained an identification rate of 96.0%. Moreover, some studies used both local and global features for writer identification [11,21].

In existing studies, there have been various writer identification methods such as SVM [22,23], distance-based [9,24], deep learning [25], and so on. For example, Nakamura et al. [9] used online text-independent data of handwritten-based Kanji characters and collected 1230 samples from 41 respondents. They extracted different kinds of features using a one-way analysis of variance (ANOVA). Moreover, the distance-based method was adopted to separate the writers and obtained an identification accuracy of 90.4%. Soma et al. [11] utilized offline handwritten-based Kanji characters for writer identification and obtained 99.0% identification accuracy using a voting method for three Kanji characters. Namboodiri and Gupta [13] used a neural network (NN) to identify the writers using online handwritten-based English characters and achieved a classification accuracy of 88.0%. Li et al. [14] proposed k-nearest neighbors (k-NN) for writer identification based on online handwritten-based English characters and obtained an identification accuracy rate of 93.6%. Wu et al. [15] employed a hidden Markov model (HMM) for the identification of writers and achieved a 95.5% identification accuracy rate. Nguyen et al. [17] also adopted the convolution neural network (CNN)-based model for writer identification and reported 93.8% identification accuracy. Rehman et al. [6] also proposed deep learning for the identification of writers using handwritten-based Chinese characters. They collected 52,800 samples and obtained 99.0% identification accuracy. Abdi et al. [4] also proposed k-NN method for writer identification using offline handwritten-based Arabic characters and achieved 90.2% identification accuracy. Nasuno and Arai [10] conducted an experiment on Japanese handwritten characters that contained 100 Japanese characters written 50 times by 100 subjects. They trained AlexNet CNN for 90 Japanese characters, tested model performance on 10 characters, and obtained an identification accuracy of 90.0%.

3. Materials and Methods

3.1. Device for Data Collection

Handwritten-based Kanji character data were collected using a pen tablet system (Cintiq Pro 16, Wacom Co., Ltd., Saitama, Japan). The pen tablet was connected to a laptop PC with Windows 10. The screen size of the pen tablet was 15.6 inches, and the resolution size was $2560\times 1440$ pixels. The coordinates of generated parameters using a pen tablet are shown in Figure 1.

3.2. Dataset Preparation

The dataset utilized in this study was collected using a pen tablet system. The dataset consisted of 33 people who wrote 72 Kanji characters 20 times each. A total of $33\times 72\times 20$ = 47,520 samples were used for this study. The handwriting data contained six pieces of information: the x-coordinate and y-coordinate, which are the positions where the characters were written; the writing pressure; the azimuth angle; the altitude; and the time taken to acquire the data.

3.3. Proposed Methodology

The goal of this study is to propose an efficient writer identification system using handwritten-based Kanji characters. The proposed model of writer identification has the following tasks: feature extraction, feature normalization, feature selection, tune the hyperparameters and train the SVM-based model, and writer identification. The block diagram of the proposed model for writer identification is depicted in Figure 2. Every step or phase is more clearly explained in the following subsections.

3.4. Feature Extraction

We collected six raw features as the x-coordinate and y-coordinate, which are the positions where the characters were written, the writing pressure, the azimuth angle, the altitude, and the time taken to acquire the data. Based on these six raw features, we computed forty features. The calculation of speed is based on [26], and peak instantaneous speed, acceleration, peak instantaneous acceleration, positive pressure change, and negative pressure change are based on [22]. The mean and standard deviation (std) of the beginning and end states of writing pressure, altitude, azimuth, speed, and acceleration were taken from previous studies [27,28,29]. An explanation of each feature, along with their descriptions and calculation formula, is given in

Table 2. List of extracted features names, their descriptions, and calculation formulas.

SN	Features	Descriptions	Calculation Formula
1	Speed mean	Mean of speed.	$\overline{s}=\frac{1}{n}{\sum }_{i=0}^n{s}_i$
2	Speed std	Std of speed.	$s\_\sigma =\sqrt{\frac{1}{n}{\sum }_{i=0}^n{{(s}_i-\overline{s})}^2}$
3	Max speed	Maximum speed.	$s\_max=Max(s_i)$
4	First speed mean	Mean of the speed of the first 10% of the whole.	$\overline{s}=\frac{1}{k}{\sum }_{i=0}^k{s}_i;(k=\frac{n}{10})$
5	First speed std	Std of the speed of the first 10% of the whole.	$s\_\sigma =\sqrt{\frac{1}{k}{\sum }_{i=0}^k{{(s}_i-\overline{s})}^2};(k=\frac{n}{10})$
6	Last speed mean	Mean speed of the last 10% of the whole.	$\overline{s\_last}=\frac{1}{n-k}{\sum }_{i=k}^n{s}_i;(k=\frac{9}{10}n)$
7	Last speed std	Std of the speed of the last 10% of the whole.	$s\_last\_\sigma =\sqrt{\frac{1}{n-k}{\sum }_{i=k}^n{{(s}_i-\overline{s})}^2};(k=\frac{9}{10}n)$
8	Piv	Maximum speed recorded at any time.	$V_{max}(V_i=\frac{L_i}{T_i})$
9	Accel. mean	Mean of accel.	$\overline{ac}=\frac{1}{n}{\sum }_{i=0}^n{ac}_i$
10	Accel. std	Std of accel.	$ac\_\sigma =\sqrt{\frac{1}{n}{\sum }_{i=0}^n{\left({ac}_i-\overline{ac}\right)}^2}$
11	Max accel.	Maximum accel.	$ac\_max=Max({ac}_i)$
12	First accel. mean	Mean of accel. of the first 10% of the whole.	$\overline{ac\_1st}=\frac{1}{k}{\sum }_{i=0}^k{ac}_i;k=\frac{n}{10}$
13	First accel. std	Std of accel. of the first 10% of the whole.	$ac\_1st\_\sigma =\sqrt{\frac{1}{k}{\sum }_{i=0}^k{\left({ac}_i-\overline{ac}\right)}^2};k=\frac{n}{10}$
14	Last accel. mean	Mean of accel. of the last 10% of the whole.	$\overline{ac\_last}=\frac{1}{n-k}{\sum }_{i=k}^n{ac}_i;k=\frac{9}{10}n$
15	Last accel. std	Std of accel. of the last 10% of the whole.	$ac\_last\_\sigma =\sqrt{\frac{1}{n-k}{\sum }_{i=0}^n{\left({ac}_i-\overline{ac}\right)}^2};(k=\frac{9}{10}n)$
16	Pia	Maximum accel. recorded at any point.	$A_{max}(A_i=\frac{V_i}{T_i})$
17	Pressure mean	Mean of pen pressure.	$\overline{p}=\frac{1}{n}{\sum }_{i=0}^n{p}_i$
18	Pressure std	Std of pen pressure.	$p\_\sigma =\sqrt{\frac{1}{n}{\sum }_{i=0}^n({p_i-\overline{p})}^2}$
19	Max pressure	Maximum of pen pressure.	$p\_max=Max(p_i)$
20	First pressure mean	Mean of pen pressure of the first 10% of the whole.	$\overline{p\_1st}=\frac{1}{n}{\sum }_{i=0}^n{p}_i;(k=\frac{n}{10})$
21	First pressure std	Std of pen pressure of the first 10% of the whole.	$p\_1st\_\sigma =\sqrt{\frac{1}{k}{\sum }_{i=0}^k({p_i-\overline{p})}^2};(k=\frac{n}{10})$
22	Last pressure mean	Mean of pen pressure of the last 10% of the whole.	$\overline{p\_last}=\frac{1}{n-k}{\sum }_{i=k}^n{p}_i;(k=\frac{9}{10}n)$
23	Last pressure std	The std of pen pressure of the last 10% of the whole.	$p\_last\_\sigma =\sqrt{\frac{1}{n-k}{\sum }_{i=k}^n({p_i-\overline{p})}^2};(k=\frac{9}{10}n)$
24	Azimuth mean	Mean of azimuth.	$\overline{az}=\frac{1}{n}{\sum }_{i=0}^n{az}_i$
25	Azimuth std	Std of azimuth.	$az\_\sigma =\sqrt{\frac{1}{n}{\sum }_{i=0}^n{{(az}_i-\overline{az})}^2}$
26	First azimuth mean	Mean of the azimuth of the first 10% of the whole.	$\overline{az\_1st}=\frac{1}{k}{\sum }_{i=0}^k{az}_i;(k=\frac{n}{10})$
27	First azimuth std	Std of the azimuth of the first 10% of the whole.	$az\_1st\_\sigma =\sqrt{\frac{1}{k}{\sum }_{i=0}^k{{(az}_i-\overline{az})}^2};(k=\frac{n}{10})$
28	Last azimuth mean	Mean of the azimuth of the last 10% of the whole.	$\overline{az\_last}=\frac{1}{n-k}{\sum }_{i=k}^n{az}_i;(k=\frac{9}{10}n)$
29	Last azimuth std	Std of the azimuth of the last 10% of the whole.	$az\_last\_\sigma =\sqrt{\frac{1}{n-k}{\sum }_{i=k}^n{{(az}_i-\overline{az})}^2};(k=\frac{9}{10}n)$
30	Altitude mean	Mean of altitude.	$\overline{alt}=\frac{1}{n}{\sum }_{i=0}^n{alt}_i$
31	Altitude std	Std of altitude.	$alt\_\sigma =\sqrt{\frac{1}{n}{\sum }_{i=0}^n({alt}_i{-\overline{alt})}^2}$
32	First altitude mean	Mean of the altitude of the first 10% of the whole.	$\overline{alt\_1st}=\frac{1}{n}{\sum }_{i=0}^k{alt}_i;(k=\frac{n}{10})$
33	First altitude std	Std of the altitude of the first 10% of the whole.	$alt\_1st\_\sigma =\sqrt{\frac{1}{k}{\sum }_{i=0}^k({alt}_i{-\overline{alt})}^2};(k=\frac{n}{10})$
34	Last altitude mean	Mean of the altitude of the last 10% of the whole.	$\overline{alt\_last}=\frac{1}{n-k}{\sum }_{i=k}^n{alt}_i;(k=\frac{9}{10}n)$
35	Last altitude std	Std of the altitude of the last 10% of the whole.	$alt\_last\_\sigma =\sqrt{\frac{1}{n-k}{\sum }_{i=k}^n({alt}_i{-\overline{alt})}^2};(k=\frac{9}{10}n)$
36	Positive pressure change mean	Mean of increase in pen pressure between two time points.	$\mathrm{Mean}(p_{i+1}-p_i/t_{i+1}-t_i$, where $p_{i+1}>pi)$
37	Positive pressure changes std	Std of increase in pen pressure between two time points.	$\mathrm{Std}(p_{i+1}-p_i/t_{i+1}-t_i$, where $p_{i+1}>pi)$
38	Max positive pressure change	Maximum increase in pen pressure between two time points.	$\mathrm{Max}(p_{i+1}-p_i/t_{i+1}-t_i$ for $p_{i+1}>pi)$
39	Negative pressure change mean	Mean decrease in pen pressure between two time points.	$\mathrm{Mean}(p_{i+1}-p_i/t_{i+1}-t_i$, where $p_{i+1}<pi)$
40	Negative pressure changes std	Std of decrease in pen pressure between two time points.	$\mathrm{Std}(p_{i+1}-p_i/t_{i+1}-t_i$, where $p_{i+1}<pi)$

.

3.5. Feature Normalization

Data normalization is a technique that minimizes redundancy and improves the efficiency of the data. Mathematically, it is defined as follows:

$$z=\frac{X-\mu }{\sigma }$$

(1)

where X is the original feature vector, $\mu$ is the mean of that feature vector, and $\sigma$ is the standard deviation. The value of z lies between 0 and 1.

3.6. Feature Selection

Feature selection is the process of removing irrelevant features and improving the model’s performance. We applied sequential forward floating selection (SFFS) to select the best combination of features. It is one kind of greedy algorithm to reduce the initial d-dimensional feature space into a k-dimensional feature subspace $(k<d)$ [30]. The pseudo-code is shown below:

• Input:$Y=\{y_1,y_2,\cdots ,y_d\}$
• Output:$X_k=\left\{x_j\right|j=1,2,\cdots ,k;x_j\in Y\}$, where $k\in (0,1,2,\cdots ,d)$
• Initialize:$X_o=\varnothing ,k=0$

Step 1: 

$x^{+}$ = argmax $J(X_k+x)$, where $x\in Y-X_k,J$ is an evaluation index and $x^{+}$ is the feature with the highest evaluation when it is chosen.

Step 2: 

$x^{-}$ = argmax $J(X_k-x)$, where $x\in X_k$ and $x^{-}$

is the feature with the best performance when the feature is deleted.

If $J(X_k-x)>J(X_k):$

$X_{k-1}=X_k-x^{-}$

$k=k-1$

Go to Step 1.

In Step 1, include the feature of the feature space that leads to the best performance increase from the feature subset. Then, go to Step 2. In Step 2, remove only a feature in the case of increasing the performance of the result subset. If k = 2 or the improvement cannot be made, go back to Step 1; else, repeat Step 2. When k = k, terminate this algorithm. In this study, we implemented SFFS using the SequentialFeatureSelctor of the “mlxtend” library [31]. We selected the combination of the best features with the highest identification accuracy.

3.7. Classification Using SVM

SVM is a supervised learning method that is used to solve problems in classification and regression. It is an effective learning method in high-dimensional spaces and when the number of dimensions is larger than the number of samples. Hyperplanes can be constructed in high-dimensional or infinite-dimensional spaces [32,33]. Intuitively, separation is achieved by the hyperplane with the largest distance to the nearest training data point in the class—the so-called margin-maximizing hyperplane. In this study, we implemented it in scikit-learn support vector classification (SVC) [34]. The main objective of SVM is to find the hyperplane in the feature space that can easily separate the classes [19], which needs to solve the following constraint problem:

$$max\alpha \sum _{i=1}^n{\alpha }_i-\frac{1}{2}\sum _{i=1}^n\sum _{j=1}^n{\alpha }_i{\alpha }_j{y}_i{y}_{j}K(x_i,x_j)$$

(2)

subject to

$$\sum _{i=1}^n{y_i}^T{\alpha }_i=1,0\le {\alpha }_i\le C,i=1,\dots ,n\forall i=1,2,3,\dots ,n$$

(3)

The final discriminate function takes the following form:

$$f(x)=\sum _{i=1}^n{\alpha }_{i}K(x_i,x_j)+b$$

(4)

where b is the bias term.

SVM can be used for both binary classification problems and multiclass classification problems [35]. For multiclass problems, there are two different approaches: (i) one-vs-one approach and (ii) one-vs-all approach. In the case of this study, we have used a one-vs-all approach.

4. Results and Discussion

In this study, we used SVM with two kernels, either linear or radial basis function (RBF), for classification. These kernels had some additional parameters called hyperparameters. We tuned these hyperparameters using the grid search method. In the current study, we performed two experiments: (i) text-dependent and (ii) text-independent. The details are explained in the following sections.

4.1. Selected Features Using SFFS

In the current study, we adopted SFFS to identify the potential features for writer identification. We selected 30 out of 40 features using SFFS. The list of selected features is presented in

Table 3. List of selected features using SFFS.

SN	Features	SN	Features	SN	Features
1	Accel. mean	11	Altitude mean	21	Altitude std
2	Azimuth mean	12	Azimuth std	22	Altitude mean first accel. mean
3	First azimuth mean	13	First pressure mean	23	First pressure std
4	First speed mean	14	First speed std	24	Last accel. mean
5	Last accel. std	15	Last altitude mean	25	Last azimuth mean
6	Last speed mean	16	Last pressure mean	26	Last pressure std
7	Last azimuth mean	17	Last speed std	27	Max accel.
8	Max pressure	18	Negative pressure change mean	28	Negative pressure changes std
9	Pressure mean	19	Positive pressure change mean	29	Positive pressure changes std
10	Pressure std	20	Speed mean	30	Speed std

.

4.2. Text-Dependent

In the text-dependent type, we applied SVM for all 72 Kanji characters. We performed five-fold cross-validation for each character and computed the identification accuracy. The identification accuracy of SVM for each character is presented in

Table 4. Identification accuracy (in %) of SVM for each character.

SN	Kanji	ACC	SN	Kanji	ACC	SN	Kanji	ACC	SN	Kanji	ACC
1	避	99.2	19	臓	97.4	37	衛	96.6	55	甘	95.9
2	担	98.6	20	岡	97.2	38	右	96.5	56	鯛	95.7
3	甫	98.6	21	湖	97.1	39	肇	96.5	57	憲	95.7
4	還	98.3	22	旬	97.1	40	委	96.3	58	響	95.4
5	虚	98.1	23	完	97.1	41	込	96.3	59	麗	95.4
6	鐵	98.1	24	崩	96.9	42	末	96.3	60	蹟	95.3
7	旭	98.0	25	車	96.9	43	凰	96.3	61	圏	95.1
8	魅	98.0	26	鞠	96.9	44	雪	96.2	62	藁	95.1
9	鳥	98.0	27	性	96.9	45	方	96.2	63	名	95.1
10	柿	97.8	28	釉	96.9	46	石	96.2	64	台	94.7
11	惹	97.7	29	亜	96.9	47	羊	96.2	65	四	94.5
12	讃	97.5	30	巣	96.9	48	曲	96.2	66	士	94.3
13	君	97.5	31	函	96.8	49	区	96.0	67	今	93.6
14	鋒	97.5	32	蘭	96.8	50	寵	96.0	68	升	93.1
15	策	97.4	33	忘	96.8	51	機	96.0	69	川	92.5
16	歯	97.4	34	渡	96.6	52	禾	96.0	70	上	92.5
17	武	97.4	35	憾	96.6	53	刊	95.9	71	入	91.2
18	費	97.4	36	滝	96.6	54	彌	95.9	72	人	84.2

. The identification accuracy of SVM per character ranged from 84.2% to 99.2%. SVM gave the highest accuracy of 99.2% for “避” and the lowest accuracy of 84.2% for “人”. Compared with the characters with high and low accuracy, the characters with many strokes had high accuracy and the characters with few strokes had low accuracy.

We also completed an experiment by taking more than one character (called connected characters) at a time and trying to show the performance of SVM for writer identification.

Table 5. Identification accuracy (in %) of connected characters.

Kanji	ACC	Cost	Gama	Kernel
避担	99.4	0.1	0.00001	Linear
避担甫	99.6	10	0.01	RBF
避担甫還	99.4	0.1	0.00001	Linear

shows the identification accuracy of SVM for two, three, and four connected characters. In order to make two, three, and four connected characters, we selected the top two (避and 担), three (避, 担, and 甫), and four characters (避, 担, 甫, and 還) from Table 4. SVM provided 99.4% and 99.6% identification accuracy for connecting two and three characters, respectively.

4.3. Text-Independent

For text-independent, we split the dataset into two sets as training and test sets. We took 70% of the dataset for training and another 30% for the test set. In this section, we used the top 5, top 10, top 15, top 20, top 25, top 30, top 35, top 40, top 45, top 50, top 55, top 60, top 65, top 70, and all 72 characters to show how the identification accuracy of SVM changes depending on whether features are selected or not over characters. Here, the person can write any combination of the top 5, top 10, etc. characters in order to make it text-independent writer identification.

Table 6. Identification accuracy (in %) of Kanji characters used for each of the five letters.

SN	Characters	SVM-WO-SFFS	SVM-W-SFFS
1	Top 5	96.2	99.0
2	Top 10	96.5	97.7
3	Top 15	96.0	97.0
4	Top 20	96.2	97.4
5	Top 25	95.6	97.1
6	Top 30	96.1	96.7
7	Top 35	95.6	96.9
8	Top 40	95.5	96.9
9	Top 45	95.5	96.3
10	Top 50	95.3	96.1
11	Top 55	94.7	95.9
12	Top 60	94.8	95.6
13	Top 65	94.2	95.3
14	Top 70	94.1	95.0
15	All 72	93.9	94.3

shows the identification accuracy of SVM without SFFS (SVM-WO-SFFS) and SVM with SFFS (SVM-W-SFFS). It was observed that SVM-WO-SFFS provided 96.2% identification accuracy for the top 5 characters, whereas SVM-W-SFFS also provided 99.0% identification accuracy. It was also observed that the identification accuracy of SVM-W-SFFS decreased (99.0% to 94.3%) when increasing the number of characters.

4.4. Comparison of Our Proposed Identification System and Similar Existing Studies

In this section, we made a comparison between our proposed writer identification system against similar existing studies. This comparative study is shown in

Table 7. Comparison of our proposed system against similar existing studies.

Authors	Data Types	Study Types	Samples	Methods	Language	Characters	Writers	ACC (%)
Nakamura et al. [9]	Online	Text-independent	1230	Based on Distance	Kanji	4	41	96.5
Soma and Arai [24]	Offline	Text-dependent	50,000	Based on Distance	Kanji	100	100	95.2
Soma et al. [11]	Offline	Text-dependent	50,000	Majority Voting	Kanji	100	100	99.0
Nguyen et al. [17]	Offline	Text-independent	2965	CNN	Kanji	100	400	93.8
Our proposed	Online	Text-independent	47,520	SVM	Kanji	72	33	99.0
		Text-dependent	47,520	SVM	Kanji	72	33	99.6

, which presents various parameters such as author names, data type (DT), study type, sample size, classification methods, language, number of characters and writers, number of selected characters and writers, and identification accuracy rate. Some existing studies were conducted for writer identification using Kanji characters as follows:

Nakamura et al. [9] collected a total of 1230 samples using pen tablets from 41 subjects (males: thirty-five; females: six). They asked each of the subjects to write only four Kanji characters five times, and repeated this six times $(5\times 6\times 41)$. They extracted 563 features and evaluated their discriminative power using one-way ANOVA and Kruskal–Wallis test. They separated the writers in feature space based on distance and achieved an identification accuracy of 90.4% for all 563 features; an identification accuracy of 96.5% was also obtained for 270 selected features.

Soma and Arai [24] also proposed a recognition system for text-dependent writer identification except using character recognition features. They obtained five features, such as start and end points of both x and y coordinates, and the angle of the stroke from each stroke. They conducted their study on 50,000 samples $(100\times 100\times 50)$, which had 100 Kanji characters, 100 subjects, and 50 samples of each character’s class for a writer. They used 5000 samples (100 writers × 50 samples) for the experiment of each character class. They used 4999 samples for the dictionary and one sample for validation. Then, Euclidian distance was used to determine the writers and achieved an identification accuracy of 99.6%, 97.0%, and 95.2% for 10 writers, 50 writers, and 100 writers, respectively. Here, the authors obtained good identification accuracy when they considered only 10 writers.

Soma et al. [11] conducted another study on the same database [24] and proposed efficient character features and a recognition system for only text-dependent writer identification. They extracted various local and global features from each character. These extracted features were fed into a majority voting approach with a leave-one cross-validation protocol for writer identification and obtained an identification accuracy rate of 99.0% for three-character classes. Nguyen et al. [17] proposed a CNN-based approach for offline text-independent writer identification. They illustrated that their proposed method obtained an identification accuracy of 99.9% for 200 characters and 100 writers; 92.8% identification accuracy for 50 characters and 100 writers; and 93.8% identification accuracy for 100 characters and 400 writers. Although the authors achieved good identification accuracy, it required writing 200, 50, or 100 characters to identify a person, which is very time-consuming and laborious.

As mentioned, the above studies were conducted for writer identification using online or offline handwritten Kanji characters based on only text-independent [9,17] or text-dependent methods [11,24] (Soma et al., 2013; Soma et al., 2014). Moreover, Nakamura et al. [9] conducted their study on only four online Kanji characters. As shown in Table 7, we used both text-dependent and text-independent writer identification methods in this work. Our experimental results demonstrated that SVM achieved the highest identification accuracy of 99.6% for only the three connected characters text-independent and 99.0% identification accuracy for only the top five characters text-dependent, which are shown in Table 7.

5. Conclusions and Future Work Direction

In this study, we proposed an efficient method for writer identification using online handwritten-based Kanji characters. The text-dependent and text-independent writer identification methods were used. First, we extracted different features from handwriting data; then, we applied SVM for writer identification. Parameter tuning was performed to increase the identification accuracy of SVM for writer identification. SVM provided the highest identification accuracy of 99.0% for the text-independent and 99.6% for the text-dependent methods. Furthermore, the characters with the highest accuracy were not simple characters with few strokes, such as “人” and “入”, but characters with distinctive strokes, such as “避” and “担”. As a result of connecting the top characters with high accuracy, we were able to efficiently obtain a high-performing system. The results of the experiment showed that some of these features, such as the relative position of the start of writing, the length of writing, and the tilt of the pen and its movement, are significant enough to identify and verify the writer. We hope that our proposed method will be helpful in various applications, such as providing evidence to a forensic expert in identifying the writer or verifying and authenticating the writer in a court of law or bank. In the future, we will extend this work by adding more subjects or users and use a deep-learning-based approach to identify writers.