PARADIGMA BARU PENDIDIKAN MATEMATIKA DAN APLIKASI ONLINE INTERNET PEMBELAJARAN


3 

How to cite: M. Athoillah, “K-Nearest Neighbor for Recognize Handwritten Arabic 

Character”, Mantik, vol. 5, no. 2, pp. 83-89, October 2019. 

K-Nearest Neighbor for Recognize Handwritten Arabic Character  
 

 

Muhammad Athoillah 

Universitas PGRI Adi Buana Surabaya, athoillah.muhammad@gmail.com 
 

doi: https://doi.org/10.15642/mantik.2019.5.2.83-89 

 

 
Abstrak: Pengenalan teks tulisan tangan adalah kemampuan sebuah sistem untuk mengenali tulisan 

tangan manusia dan mengubahnya menjadi teks digital. Pengenalan teks tulisan tangan adalah 
bagian dari masalah klasifikasi, sehingga algoritma klasifikasi seperti Nearest Neighbor (NN) 

diperlukan untuk menyelesaikannya. Algoritma NN adalah algoritma yang sederhana namun 
memberikan hasil yang baik. Berbeda dengan algoritma lain yang biasanya ditentukan oleh beberapa 

kelas hipotesis, Algoritma NN menemukan label pada titik uji tanpa mencari prediktor dalam 
beberapa kelas fungsi yang telah ditentukan. Bahasa Arab adalah salah satu bahasa terpenting di 

dunia. Mengenali karakter Bahasa Arab sangat menarik untuk dijadikan bahan kajian, tidak hanya 
karena merupakan bahasa utama yang digunakan dalam agama Islam tetapi juga karena jumlah 

penelitian tulisan tangan Arab yang ada masih jauh jumlahnya bila dibandingkan dengan penelitian 
pengenalan tulisan Latin atau Cina. Berdasarkan latar belakang tersebut, penelitian ini membangun 

sebuah sistem untuk mengenali Karakter tulisan Arab dari sebuah citra menggunakan algoritma NN. 
Hasil penelitian menunjukkan bahwa metode yang diusulkan dapat mengenali karakter dengan 

sangat baik ditunjukkan melalui rata-rata presisi, recall dan akurasi yang tinggi. 
 

Kata kunci: Klasifikasi, Karakter Bahasa Arab, Nearest Neighbor, Pengenalan Teks, Tulisan 

Tangan 

 

 
Abstract: Handwritten text recognition is the ability of a system to recognize human handwritten 

and convert it into digital text. Handwritten text recognition is a form of classification problem, so 

a classification algorithm such as Nearest Neighbor (NN) is needed to solve it. NN algorithms is a 

simple algorithm yet provide a good result. In contrast with other algorithms that usually determined 

by some hypothesis class, NN Algorithm finds out a label on any test point without searching for a 

predictor within some predefined class of functions. Arabic is one of the most important languages 

in the world. Recognizing Arabic character is very interesting research, not only it is a primary 

language that used in Islam but also because the number of this research is still far behind the number 

of recognizing handwritten Latin or Chinese research. Due to that's the background, this framework 

built a system to recognize handwritten Arabic Character from an image dataset using the NN 

algorithm. The result showed that the proposed method could recognize the characters very well 

confirmed by its average of precision, recall and accuracy. 

 

Keywords: Arabic Character, Classification, Handwritten, Nearest Neighbor, Text Recognition 

 

  

Jurnal Matematika MANTIK 

Volume 5, Nomor 2, October 2019, pp. 83-89 

 

ISSN: 2527-3159 (print) 2527-3167 (online) 

http://u.lipi.go.id/1458103791
http://u.lipi.go.id/1457054096


Jurnal Matematika MANTIK 

Vol. 5, No. 2, October 2019, pp. 83-89 

84 

1. Introduction  

Handwritten text recognition is the ability of a system to recognize human 

handwritten and convert it into digital text, this study was first conducted by T.L 

Diamond in 1958 [1]. Recently, the study of handwritten text recognition has been 

growing rapidly, the huge number of these researches is proved by the number of 

its publication such as Sadkhan et al. in 2018 who recognized handwritten based on 

ANN and wavelet Transformation [2], Iqbal and Zafar studied about Offline 

Handwritten Quranic Text Recognition in 2019 [3], Nguyen et al. who recognized 

Japanese handwritten using Recurrent Neural Network in 2018 [4] and much more. 

Handwritten recognition is divided into two categories, namely on-line and off-line 

recognition. The on-line recognition is a system that has the ability to recognize 

humanly handwritten directly from smartphones, personal digital assistant (PDA) 

and others similar device, whilst the off-line recognition is a system that needs to 

scan any images first to recognize the human handwritten[5]. 

Arabic is one of the most important languages in the world. Since it is a primary 

language used in Islam, recognizing Arabic character had become a very interesting 

task. Study of Arabic text recognition was started by A. Nazif in 1975. In his 

master's thesis, he built a system to recognize Arabic characters based on extracting 

strokes called radicals and their positions. He used correlation among the templates 

of the character image and the radicals, a segmentation phase was included in the 

cursive text segment [6]. However, the number of recognizing handwritten Arabic 

research is still far behind the number of recognizing handwritten Latin or Chinese 

research. This fact is based on the lack of Arabic digital dictionaries and 

programming tools or the lack of public databases of handwritten Arabic characters 

that found nowadays[5].  

Handwritten text recognition is a form of classification problem, so a 

classification algorithm is needed to solve it, one of that algorithm is the Nearest 

Neighbor (NN). Nearest Neighbor algorithms are simple algorithm but give a good 

result. The idea of this method is to memorize the training data and predict the label 

of any new instance on the basis of labels from its closest neighbours in the training 

data. In contrast with other algorithms that usually determined by some hypothesis 

class, Nearest Neighbor Algorithm finds out a label on any test point without 

searching for a predictor within some predefined class of functions [7]. By default, 

the most function that is used to measure the distance between the data (neighbour) 

is Euclidean. Moreover, the parameter 𝑘 is added which decided how many 
neighbours will be chosen for kNN algorithm. The suitable choice of 𝑘 has a 
significant impact on the result of the kNN algorithm[8][9]. 

Due to that's the background, this framework built a system to recognize 

handwritten Arabic Character scanned from an image dataset using k-Nearest 

Neighbor (kNN) algorithm. The framework used MADBAse dataset provided by 

The American University in Cairo. The result of this work was presented through 

the values of its Precision, Recall and Accuracy while cross-validation method was 

used to validate the results. 

 

2. Related Work 

K-Nearest Neighbor (K-NN) was first introduced in 1951 and 1952 by Fix and 

Hodges [10][11] and then further study conducted by  Cover and Hart in 1967 [12]. 



M. Athoillah 

K-Nearest Neighbor for Recognize Handwritten Arabic Character 

85 

The algorithm was inspired by the hypothesis that "things that look alike must be 

alike" (Cover and Hart). Different with other classification algorithms, KNN is a 

classification method based on data that were located closest to the object without 

searching for a predictor within some predefined class of functions to find out the 

label on test point[7].  

 
Figure 1. Illustration of K-NN Algorithm 

 

Assume that dataset domain 𝑋 is given with metric function 𝑝, then the function 
that returns the distance between two elements of 𝑋 written as 𝑝: 𝑋 × 𝑋 → ℝ. For 
instance, if 𝑋 =  ℝ𝑑 then 𝑝 can be Euclidean distance 

𝑝(𝑥, 𝑥′) = ‖𝑥 − 𝑥′‖ =  ∑(𝑥𝑖 − 𝑥𝑖
′)2

𝑑

𝑖=1

 

Given a sequence of training data as 𝑆 = (𝑥1, 𝑦1), (𝑥2, 𝑦2), … ,( 𝑥𝑚, 𝑦𝑚). For 
each 𝑥 ∈ 𝑋, let 𝜋1(𝑥), 𝜋2(𝑥), …  , 𝜋𝑚(𝑥) be a reordering of {1,2, …  , 𝑚} according 
to their distance to 𝑥, 𝑝(𝑥, 𝑥𝑖 ). Then, for all 𝑖 < 𝑚, 

𝑝(𝑥, 𝑥𝜋𝑖(𝑥)) ≤ 𝑝(𝑥, 𝑥𝜋𝑖+1(𝑥)) 

For a number 𝑘, the binary classification of 𝑘 −NN rule is defined as follows: 
 

• Input: sample of training 𝑆 = (𝑥1, 𝑦1), (𝑥2, 𝑦2), … ,( 𝑥𝑚 , 𝑦𝑚) 
• Output: for every point 𝑥 ∈ 𝑋, return the majority label within {𝑦𝜋𝑖(𝑥) ; 𝑖 ≤ 𝑘} 

 

If 𝑘 = 1, then the 1-NN rule: 
ℎ𝑠 (𝑥) = 𝑦𝜋1(𝑥)  

 

Since the output depends on the number of 𝑘, therefore the appropriate of 𝑘 has a 
significant impact on the result of the kNN algorithm. 

 

3. Implementation 

This framework used a dataset from The American University in Cairo called 

MADBAse which contain images of Arabic handwritten Numeral Character from 

zero to nine-digit, the images used in this experiment amount to 500 in each object, 

which means there are 5000 images used in total.  



Jurnal Matematika MANTIK 

Vol. 5, No. 2, October 2019, pp. 83-89 

86 

 
Figure 2. Sample of Dataset 

 

The Image dataset was divided in two-part that was dataset for training process 

which contains 90 percent of the total image, whilst the rest 10 percent was used 

for the testing process. This experiment used histogram of the images as input, the 

histogram was chosen because the colour features which distinguish each image is 

generally represented by the colour histogram[13]. K-Nearest Neighbor is a simple 

algorithm, the process of training for this algorithm only contain storing feature 

vectors and labels of the images, while in classification process, the unlabeled query 

point is assigned to the label of its k nearest neighbours. When used K-NN, the 

object was classified based on majority labels among the k nearest neighbours. If 

k=1, the object was classified as the class from the object nearest to it. If there are 

only two classes, k must be an odd integer. However, there was still bond when k is 

an odd integer in case performing multiclass classification. In addition, Euclidean 

distance function was used to measure the distance between data.  

 

4. Result and Discussion 

In this section, the performance of the system to recognize the Arabic 

handwritten is delivered by computing its precision, recall and accuracy. Since the 

result of the K-NN algorithm is greatly influenced by the value of 𝑘, this framework 
measured all the performance used 6 sorts of 𝑘, namely 1, 3, 5,10 15 and 20. In 
addition, K-fold cross-validation was applied to validate the result of the system, 

this process was executed by dividing data into ten batches (𝑘 = 10), which are nine 
batches for training and the rest batch for testing, then the experiment was repeated 

𝑘 times under the condition that training data and testing data were always different 
during each process. The results are presented in the following table: 

 
Table 1.  The Average Result of System with 𝑘 = 1 (%) 

Class/Number Precision Recall Accuracy 

Zero 93,57 91,00 98,42 

One 92,16 94,20 98,58 

Two 22,34 20,40 84,68 

Three 13,20 15,60 81,52 

Four 14,89 15,40 83,02 

Five 72,44 68,20 94,06 

Six 40,31 25,60 88,72 

Seven 17,22 18,60 82,70 

Eight 20,79 26,40 82,50 

Nine 24,27 21,60 85,20 

Average off All 42,99 41,71 88,24 

 

 



M. Athoillah 

K-Nearest Neighbor for Recognize Handwritten Arabic Character 

87 

Table 2.  The Average Result of System with 𝑘 = 3 (%) 
Class/Number Precision Recall Accuracy 

Zero 94,81 91,00 98,56 

One 92,47 96,20 98,80 

Two 23,22 20,80 84,80 

Three 14,04 15,60 82,22 

Four 17,38 17,80 83,46 

Five 72,80 74,40 94,50 

Six 44,09 27,60 89,18 

Seven 18,31 19,20 83,28 

Eight 21,24 27,60 82,42 

Nine 26,53 23,40 85,50 

Average off All 44,26 43,36 88,58 

 

 

Table 3.  The Average Result of System with 𝑘 = 5 (%) 
Class/Number Precision Recall Accuracy 

Zero 94,63 91,40 98,58 

One 92,41 96,00 98,76 

Two 25,82 24,00 85,36 

Three 14,29 16,20 82,22 

Four 19,07 18,80 83,78 

Five 74,24 76,00 94,82 

Six 50,80 32,60 89,96 

Seven 20,45 19,80 83,94 

Eight 20,06 25,80 82,16 

Nine 28,45 26,00 85,74 

Average off All 45,75 44,51 88,84 

 

 

Table 4.  The Average Result of System with 𝑘 = 10 (%) 
Class/Number Precision Recall Accuracy 

Zero 95,35 91,20 98,64 

One 92,12 95,80 98,70 

Two 28,47 26,80 85,50 

Three 13,90 14,80 82,58 

Four 21,49 21,60 84,20 

Five 74,31 79,80 95,00 

Six 53,01 35,80 90,08 

Seven 23,35 20,80 85,10 

Eight 21,66 28,80 82,60 

Nine 32,77 29,40 86,56 

Average off All 47,07 46,16 89,16 

 

Table 5.  The Average Result of System with 𝑘 = 15 (%) 
Class/Number Precision Recall Accuracy 

Zero 95,30 90,80 98,60 

One 92,74 95,20 98,70 

Two 31,61 29,00 86,00 

Three 16,07 16,40 83,30 

Four 22,90 22,80 84,44 

Five 72,74 80,40 94,84 

Six 51,72 39,80 89,86 

Seven 23,58 19,20 85,30 



Jurnal Matematika MANTIK 

Vol. 5, No. 2, October 2019, pp. 83-89 

88 

Class/Number Precision Recall Accuracy 

Eight 22,22 28,60 82,64 

Nine 34,12 30,60 86,88 

Average off All 47,65 46,91 89,30 

 

 

Table 6.  The Average Result of System with 𝑘 = 20 (%) 
Class/Number Precision Recall Accuracy 

Zero 95,79 89,20 98,50 

One 93,41 94,60 98,72 

Two 34,07 30,60 86,10 

Three 17,63 17,40 83,32 

Four 22,78 23,20 84,54 

Five 72,12 80,60 94,74 

Six 49,60 41,60 89,64 

Seven 25,15 18,20 86,06 

Eight 23,55 28,80 82,92 

Nine 34,51 33,00 86,90 

Average off All 48,23 47,13 89,39 

 

The result showed from Table 1 until Table 6 that K-NN could recognize the 

handwritten very well, proved by the lowest average of accuracy is 88,24% while 

the lowest average of its precision and recall are 42,99% and 41,71%. The highest 

performance belongs to the recognition of “Number one” and “Number zero”, this 

is because the shapes of them are the simplest and the most different from the other 

numbers which are only a straight vertical line for “Number one” and only a dot for 

“Number zero”. Otherwise, the lowest performance belongs to “Number Three” 

with the highest point of precision is 17,63%, 17,40% for recall point and 83,32% 

for its accuracy, that is because “Number Three” has a shape very similar to 

“Number Two” so those characters are often misclassified. This framework results 

also showed that the higher value of 𝑘 provided the higher value of its accuracy, 
precision and recall. Nevertheless, it cannot be concluded that the higher value of 

𝑘 is always better because there is no clear evidence or definite guidance to 
determine the best value of 𝑘 to get the optimal result. In general, a higher value of 
𝑘 makes it less sensitive to noise and cause smoother boundaries. As a result, it is 
almost impossible to choose the same best value of 𝑘 for different applications[13].  
 

5. Conclusion 

This framework built a system that could recognize handwritten Arabic 

character using K-Nearest Neighbour (K-NN). Database from The American 

University in Cairo called MADBase was used in this framework that contains 

Arabic handwritten Numeral Character Images from zero to nine-digit with a total 

of 5000 images. Since 𝑘 values have a lot of effect on the results, then this 
framework measured all the performance used six types of 𝑘 values, i.e. 1, 3, 5, 10, 
15 and 20. Relating to validation, K-fold cross-validation was applied to validate 

the result by dividing all data into ten batches then all of these batches were used 

for train and test process alternately. The result not only showed that the proposed 

method could recognize the characters very well confirmed by its average of 

precision, recall and accuracy but also showed that the higher 𝑘 values produced 



M. Athoillah 

K-Nearest Neighbor for Recognize Handwritten Arabic Character 

89 

the higher performance. Even so, it cannot be concluded that the higher 𝑘 values 
always present a better result because each data has its own characteristics. 

 

References 

 

[1] T. L. Dimond, “Devices for reading handwritten characters,” in Papers and 

discussions presented at the December 9-13, 1957, eastern joint computer 

conference: Computers with deadlines to meet, 1957, pp. 232–237. 

[2] S. B. Sadkhan and S. F. Jawad-SMIEEE, “Handwritten Recognition based 

on Hybrid ANN and Wavelet Transformation,” in 2018 Al-Mansour 

International Conference on New Trends in Computing, Communication, 

and Information Technology (NTCCIT), 2018, pp. 76–80. 

[3] A. Iqbal and A. Zafar, “Offline Handwritten Quranic Text Recognition: A 

Research Perspective,” in 2019 Amity International Conference on Artificial 

Intelligence (AICAI), 2019, pp. 125–128. 

[4] H. T. Nguyen, C. T. Nguyen, and M. Nakagawa, “Online Japanese 
Handwriting Recognizers using Recurrent Neural Networks,” in 2018 16th 

International Conference on Frontiers in Handwriting Recognition 

(ICFHR), 2018, pp. 435–440. 

[5] M. Shatnawi, “Off-line handwritten Arabic character recognition: a survey,” 

in Proceedings of the international conference on image processing, 

computer vision, and pattern recognition (IPCV), 2015, p. 52. 

[6] B. Al-Badr and S. A. Mahmoud, “Survey and bibliography of Arabic optical 

text recognition,” Signal Processing, vol. 41, no. 1, pp. 49–77, 1995. 

[7] S. Shalev-Shwartz and S. Ben-David, Understanding machine learning: 

From theory to algorithms. Cambridge university press, 2014. 

[8] Z. Zhang, “Introduction to machine learning: k-nearest neighbors,” Ann. 

Transl. Med., vol. 4, no. 11, 2016. 

[9] F. Amin, “Identifikasi Citra Daging Ayam Berformalin Menggunakan 

Metode Fitur Tekstur dan K-Nearest Neighbor (K-NN)”, mantik, vol. 4, no. 

1, pp. 68-74, May 2018. 

[10] E. Fix and J. L. Hodges Jr, “Discriminatory analysis-nonparametric 

discrimination: consistency properties,” California Univ Berkeley, 1951. 

[11] E. Fix and J. L. Hodges Jr, “Discriminatory analysis-nonparametric 

discrimination: Small sample performance,” California Univ Berkeley, 

1952. 

[12] T. M. Cover and P. Hart, “Nearest neighbor pattern classification,” IEEE 

Trans. Inf. theory, vol. 13, no. 1, pp. 21–27, 1967. 

[13] B. Lei, E.-L. Tan, S. Chen, D. Ni, and T. Wang, “Saliency-driven image 

classification method based on histogram mining and image score,” Pattern 

Recognit., vol. 48, no. 8, pp. 2567–2580, 2015. 

[14] C.-M. Ma, W.-S. Yang, and B.-W. Cheng, “How the parameters of k-nearest 

neighbor algorithm impact on the best classification accuracy: In case of 

parkinson dataset,” J. Appl. Sci., vol. 14, no. 2, pp. 171–176, 2014.