LABORATORIUM PSIKOLOGI LANJUT

LAPORAN PRAKTIKUM PSIKOLOGI KOGNITIF

(Feature Detection)

DISUSUN OLEH :

Nama : Apipah

NPM : 10518983

Kelas : 3PA08

Tutor : Puteri Saleha Rahmatillah

FAKULTAS PSIKOLOGI

UNIVERSITAS GUNADARMA

2021
 I. TUJUAN
Tujuan dari praktikum ini adalah untuk mendeteksi fitur-fitur dari
sebuah stimulus diantara stimulus lainnya dengan melakukan berbagai
percobaan.
II. POINT OF VIEW
Praktikum ini penting dilakukan untuk membantu Praktikan lebih
mengerti tentang materi feature detection dan diajarkan untuk
menggunakan stimulus-stimulus untuk mendeteksi target dengan mencari
suatu stimulus menggunakan fitur-fitur khasnya.
III. TEORI (FEATURE DETECTION)
A. Landasan Teori
Menurut John G. Benjafield (1997), feature detection merupakan
sebuah alternatif untuk teori template yang memungkinkan bahwa kita
mengidentifikasi pola berdasarkan fiturnya. Fitur adalah atribut, atau
properti, seperti ukuran, warna, bentuk, dan sebagainya.
Menurut Goldstein (2008), feature detection adalah neuron yang
merespons kepada fitur-fitur yang spesifik yang dianalisis dari
orientasi, ukuran dan seberapa kompleks fitur-fitur tersebut dalam suatu
lingkungan.
Menurut Neisser (dalam Grahame, 2001), feature detection adalah
fitur dasar dari suatu adegan secara otomatis diproses dengan tidak
memerlukan perhatian sadar dan parallel memungkinkan beberapa fitur
dasar tetapi terpisah untuk dianalisis sekaligus.
Jadi, feature detection adalah bagaimana individu menyaring
informasi dengan menganalisis fitur-fitur hingga akhirnya dapat
menentukan stimulus tersebut. Misalnya adalah saat kita mencari tokoh
kartun Spongebob diantara tokoh-tokoh kartun Minions, kita dapat
menyebutkan yang mana tokoh kartun Spongebob karena kita tau fitur-
fitur dan ciri-ciri dari tokoh kartun Spongebob.
Menurut Oliver Selfridge (1959), pandemonium yaitu sebuah
paradigma untuk belajar untuk simposium pada mekanisasi proses
pemikiran. Dimana pemerintah pusat menghipotesis bahwa surat-surat
diidentifikasi melalui fitur-fitur komponen.
Hal yang paling dikenal pada teori feature detection adalah
pandemonium (Selfridge, 1959; Norman, 1972, dalam Friedenberg dan
Silverman, 2006), Ini diambil dari nama mental kecil “demons” yang
mewakili pemrosesan suatu unit. “Demons” ini akan “berteriak” ketika
merekognisi prosesnya, misalnya sebagai contoh huruf R. Stimulus atau
huruf R, diwakili sebagai image demon yang mempertahankan
keseluruhan dari huruf tersebut. Kemudian ada satu feature demon
untuk setiap kemungkinan ciri stimulus. Feature demon berteriak jika
melihat cirinya di suatu gambaran. Huruf R mempunyai satu garis tegak
lurus dan satu lingkaran menghadap ke kanan di bagian atas, jadi huruf
R mempunyai dua feature demon. Tahap selanjutnya yaitu cognitive
demon. Jika para demon mengetahui ada yang sesuai dengan ciri dari
huruf R, maka para demon akan berteriak dan cognitive demon akan
berteriak paling keras. Dan yang terakhir yaitu decision demon.
Decision demon akan mendengarkan teriakan paling keras dari
cognitive demon sebagai huruf yang dikenali.
Menurut Selfridge (dalam Grahame, 2001), pandemonium dibagi
kedalam empat tahap pemrosesan, yaitu:
1. Image Demon untuk mencatat gambaran atau citra eksternal.
2. Feature Demon untuk melihat fitur-fitur khusus pada pola.
3. Cognitive Demon untuk mengenali suatu pola, apabila menemukan
fitur yang cocok maka demon tersebut berteriak.
4. Decision Demon untuk mendengarkan hasil pandemonium dari
cognitive demon, lalu decision demon memilih teriakan dari
cognitive demon yang berteriak paling keras sebagai pola yang
paling besar kemungkinan terjadinya.

Jadi, pandemonium merupakan mekanisme proses terjadinya

rekognisi yang diidentifikasi melalui fitur-fitur komponen yang
diindera oleh individu. Pandemonium dibagi menjadi empat tahap
 pemrosesan, yaitu image demon, feature demon, cognitive demon, dan
decision demon.

B. Jurnal Terkait
(Recognition of One-Stroke Symbols By Humans and Computers)
Jurnal yang berjudul “Recognition of One-Stroke Symbols by
Humans and Computers” ini berisi tentang penelitian mengenai
perbedaan kemampuan antara manusia dan komputer dalam mengenali
huruf yang terdistorsi. Para Peneliti menganalisis bagaimana manusia
dapat mengenali karakter atau huruf yang terdistorsi dan terpotong,
dalam penelitian tersebut dipilih empat karakter dari alphabet yaitu L,
W, s, dan e dengan memutar dan menambahkan beberapa goresan pada
huruf tersebut. Dalam penelitian, menggunakan empat tahap
pemrosesan pandemonium sebagai karakteristik bentuk karakter untuk
menganalisis hasil percobaan. Huruf alphabet dicatat di retina sebagai
image demon dan dibandingkan dengan feature demon yang mewakili
fitur tertentu seperti garis vertikal, garis horizontal, garis miring, sudut
dan sebagainya. Selanjutnya cognitive demon mencari fitur yang paling
dekat dengan kombinasi ciri huruf tertentu yang selanjutnya akan
dipilih dan itu disebut decision demon kemudian ia mengenali karakter
tertentu. Dalam percobaan, peneliti hanya memeriksa apakah karakter
yang terdistorsi dikenali sebagai aslinya karakter atau tidak. Dari hasil
penelitian didapatkan bahwa manusia tetap bisa mengenali gambar
karakter yang terdistorsi dan terpotong dengan probabilitas tinggi,
sedangkan program komputer tidak bisa.
Pada penelitian dalam jurnal terjadi proses feature detection dimana
manusia dan program komputer menyaring informasi dengan
menganalisis fitur-fitur hingga akhirnya dapat menentukan stimulus
tersebut. Namun karena pada penelitian sengaja dibuat adanya distorsi
pada huruf maka terlihat perbedaan kemampuan antara manusia yang
masih dapat mengenali huruf dengan probabilitas tinggi walaupun
 adanya distorsi dan kemampuan program komputer yang tidak dapat
mengenali huruf karena adanya distorsi tersebut.
IV. PELAKSANAAN DAN HASIL PERCOBAAN
A. Langkah-langkah
1. Masuk ke dalam web cogscidemos.swarthmore.edu
2. Lalu, klik pada bagian Vision Search.
3. Setelah itu, klik bagian Try it!
4. Selanjutnya, pilih bagian Full Experiment dan mengisi jumlah
distractors dengan angka 2, lalu isi distractor 1 dengan huruf q dan
distractor 2 dengan huruf b, kemudian targetnya diisi dengan huruf
p.
5. Lalu, klik bagian Start Experiment.
6. Setelah itu, akan muncul instruksi dimana tombol spasi ditekan
untuk melanjutkan ke soal berikutnya, tombol huruf “F” ditekan
apabila melihat target, dan tombol huruf “J” ditekan apabila tidak
melihat target.
7. Kemudian, apabila telah selesai melakukan percobaan Praktikan
meng-screenshot hasil dari percobaan yaitu berupa grafik dan tabel.
B. Hasil
Berdasarkan praktikum feature detection yang dilaksanakan pada
hari Kamis, 08 April 2021 berikut hasil yang didapatkan pada saat
praktikum berlangsung.
C. Pembahasan
Pada saat melakukan percobaan terjadi proses feature detection,
dimana Praktikan menyaring informasi dengan menganalisis fitur-fitur
yang khas hingga akhirnya dapat menentukan stimulus yang menjadi
target dalam percobaan. Dalam percobaan yang telah dilakukan juga
terjadi proses pandemonium, dimana pada awalnya Praktikan melihat
keseluruhan gambarnya (image demon), kemudian mencari target
sesuai fitur-fitur khususnya (feature demon), selanjutnya
memastikannya dengan membandingkan dengan huruf yang lain
(cognitive demon), lalu memutuskan apakah targetnya ada atau tidak
(decision demon).
Berdasarkan dari hasil percobaan yang telah Praktikan lakukan,
jumlah waktu yang dibutuhkan bertambah untuk mencari target ketika
 item distractor ditambahkan. Dalam melakukan percobaan, jumlah item
distractor di layar pada setiap uji coba bervariasi antara 4 dan 28.
Biasanya, waktu respons untuk mencari target akan lebih lama karena
lebih banyak item distractor yang ditambahkan, tetapi jumlah waktu
respons per item merupakan ukuran yang cukup baik.
Sama halnya dengan manusia, program komputer juga dapat
melakukan proses feature detection. Namun ketika karakter atau huruf
terdistorsi program komputer cenderung mengalami kesulitan atau
bahkan tidak mengenali karakter atau huruf tersebut. Berbeda dengan
kemampuan alami manusia, dimana ketika karakter sengaja dibuat
terdistorsi atau ketika harus mencari target diantara banyaknya
distractor, manusia cenderung masih mampu mengenali dan melacak
stimulus yang menjadi target.
V. KESIMPULAN
Feature detection adalah bagaimana individu menyaring informasi
dengan menganalisis fitur-fitur hingga akhirnya dapat menentukan stimulus
tersebut. Hal yang paling dikenal pada teori feature detection adalah
pandemonium. Pandemonium merupakan mekanisme proses terjadinya
rekognisi yang diidentifikasi melalui fitur-fitur komponen yang diindera
oleh individu. Pandemonium dibagi menjadi empat tahap pemrosesan, yaitu
image demon, feature demon, cognitive demon, dan decision demon.

Daftar Pustaka

Benjafield, G. J. (1997). Cognition. Amerika Serikat: Prentice Hall.

Goldstein, E. B. (2008). Cognitive Psychology: Connecting Mind, Research and

Everyday Experience. Amerika Serikat: Wadsworth Publishing.

Hill, G. (2001). A Level Psychology Through Diagrams. Inggris: Oxford University

Press.

Friedenberg, J., & Silverman, G. (2006). Cognitive Science: An Introduction to the

Study of Mind. London: Sage Publications.
Selfridge, O. G. (1959). Pandemonium: A Paradigm for Learning, In Proceedings
of the Symposium on Mechanisation of Thought Processes. Teddington:
National Physical Laboratory.

Takaya, M., Kato, H., Komatsubara, T., Watanabe, Y., & Yamamura, A. (2013).
Recognition of one-stroke symbols by humans and computers. Social and
Behavioral Sciences, 97, 666-674. Doi: 10.1016/j.sbspro.2013.10.286
 Available online at www.sciencedirect.com

ScienceDirect
Procedia - Social and Behavioral Sciences 97 (2013) 666 – 674

The 9th International Conference on Cognitive Science

Recognition of one-stroke symbols by humans and computers

Mayumi Takaya, Hikari Kato, Takeshi Komatsubara, Yusuke Watanabe,
Akihiro Yamamura*
Department of C
Computer Science and Engineering
g, Akita University,
y
1-1 Tegata Gakuen-machi, Akita 010-8502, Japan

Abstract

We study recognition ability of humans and line trace programs in order to construct a reverse Turing test based on one-stroke
symbol traces. Our aim is to analyze the difference between humans and computers from the point of view of recognizing
distorted and nicked character images. We carry out experiments to show it is tractable for humans to recognize and trace a one-
stroke character or symbol, but intractable for computer programs.

©
© 2013 TheAuthors.
2013 The Authors.Published
Publishedbyby Elsevier
Elsevier Ltd.Ltd.
Open access under CC BY-NC-ND license.
Selection and/orpeer-review
Selection and/or peer-review under
under responsibility
responsibility of Universiti
of the the Universiti Malaysia
Malaysia Sarawak
Sarawak.

Keywords: recognition; one-stroke symbol; reverse Turing test; CAPTCHA; pandemonium architecture; touchscreen; line trace simulator

1. Introduction

Humans distinguish machines and humans in the Turing test [8], while machines distinguish in the reverse
Turing test. A CAPTCHA (Completely Automated Public Turing test to tell Computers and Humans Apart) is one
of reverse Turing tests that distinguish an access by a computer program such as a crawler from an access by
humans using the difference of shape recognition ability between humans and computer programs [1, 2]. Computer
programs are used to automatically access to network services and acquire a large amount of web mail accounts
aiming at illegal network utilization such as sending spam mails. A CAPTCHA is a web security technique to detect
such illegal accesses to network services. When facing the existing CAPTCHA, a human recognizes a word, maybe
a meaningless sequence of characters, shown on a PC display and responds by typing it through the keyboard. The
character image is usually distorted in some fashion, and computer programs like an OCR cannot recognize the
original characters. Therefore the CAPTCHA is one of the applications of cognitive science to information security.
Note that reCAPTCHA, which digitalizes old printed material by asking users to decipher scanned words from
books [3], is another application of cognitive science to information communication technology. However, it has not
been completely understood difference of recognition ability of humans and computers. In this paper, we focus on
the technique proposed by Tsuruta, Takaya and Yamamura [7], in which stroke order plays a significant role in
addition to the character recognition, and analyze the difference of recognition between humans and computers.
Smartphones are convenient and inevitable in the information communication society and the development of
cloud computing promotes the spread of smartphones. Thus smartphones affect the use of the Internet. Accesses

* Akihiro Yamamura. Tel.: +81-18-889-2799; fax: +81-18-837-0408.

E-mail address: yamamura@ie.akita-u.ac.jp

1877-0428 © 2013 The Authors. Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
Selection and/or peer-review under responsibility of the Universiti Malaysia Sarawak.
doi:10.1016/j.sbspro.2013.10.286
 Mayumi Takaya et al. / Procedia - Social and Behavioral Sciences 97 (2013) 666 – 674 667

from smartphones to the Internet services are rapidly increasing, and many web sites have begun to correspond to
smartphone users. There are many differences between smartphones and the past computer models from the point of
view of human-computer interaction. For smartphones, a touchscreen is used to input data, and the display of a
smartphone is comparatively small. When we use a smartphone and face a CAPTCHA, both the challenge image
and the virtual keyboard are displayed, and we have to type the word in the challenge image. However, the virtual
keyboard occupies almost half of the display and so the CAPTCHA image must be small. This forces the length of
the word shorter and the security level lower. To solve this problem, a new CAPTCHA suitable for smartphones
using embodied knowledge of humans is proposed by [7]. As other standard authentication protocols, an Internet
web server sends a CAPTCHA challenge and the user has to respond it in a correct way; the server sends a challenge
image that includes a character (or a symbol), and then the user is requested to trace the character by a finger tip and
sends back the data representing a one-stroke sketch as a response to the challenge. The smartphone interprets the
date inputted as an ordered series of coordinates of the display in which the stroke order is represented by time
series. The web server receives the ordered series of coordinates and checks whether or not it is appropriate as a data
inputted by a human using implicitly embodied knowledge. If the data received is determined to be an output of a
computer program then the access is rejected. Thus the correctness of stroke order is vital in the technique and
therefore it is plausible to employ a line trace simulator to attack the CAPTCHA.

2. Characteristics of character images

We analyze how humans recognize and trace distorted and nicked one-stroke characters. We choose four
characters from the alphabet that can be written by a single stroke. Then we distort and nick the characters, however,
we guess humans can recognize these characters with high probability. We then study the difference between
recognition ability.
In our experiment, we choose the characters L , W , s , e can be written by one stroke, as the
basic image. Then we transform these four characters by rotating and adding several nicks in character image. An
image is given as a 300 300 pixel bitmap image and we set the color of the character the black (RGB values are all
0) and background the white (RGB values are all 255). We did not take the effect of colors into consideration in this
paper although colors may affect recognition ability as well. We show the characteristics of characters in Table 1.
The basic character images are given in Fig 1, and we name them L00, W00, s00, and e00, respectively, where 00
means that no transformation is given.
We employ the so-called Pandemonium architecture as characteristics of shape of characters to analyse our
experimental results. Selfridge introduced Pandemonium architecture to explain the image constancy phenomena [5,
6]. Pandemonium architecture is composed of several groups of demons working independently to process the
visual stimulus. It consists of image, feature, cognitive and decision demons. Each group of demons is assigned to a
specific stage in recognition. According to the architecture, the alphabetical character is recorded in the retina as an
image demon and is compared with feature demons, which are representing a specific feature such as vertical lines,
horizontal lines, oblique lines, angles and so on. Next the cognitive demon closest to a combination of feature
demons is selected and it calls decision demons. Then it recognizes a specific character. Therefore, an image is first
perceived in its part before the whole because each feature demon corresponds to image demons. The reader is
referred to [4] for detailed information.

Table 1. The characteristics of

Character L W s e
Characteristics of shape vertical line, 4 oblique lines, continuous curve horizontal line,
horizontal line, 2 acute angles continuous curve,
right angle right angle
Height [pixel] 255 185 190 190
Width [pixel] 130 255 120 150
Name L00 W00 s00 e00
668 Mayumi Takaya et al. / Procedia - Social and Behavioral Sciences 97 (2013) 666 – 674

(a) L00 (b) W00 (c) s00 (d) e00

Fig. 1. Basic character images

The feature demons are classified into (1) the vertical line, (2) horizontal line, (3) oblique line, (4) right angle, (5)
acute angle, (6) continuous curve and (7) discontinuous curve. In our experiment, the character shape is
corresponding to one of the feature demons. Note that the images L00, W00, s00, and e00 possess all features but
the discontinuous curve. We also note that the images L00, W00, and e00 have either acute or right angle feature.
Therefore, s00 is the only image without acute or right angle feature.

3. Recognition by humans

3.1. Experimental method

The basic character images L00, W00, s00, and e00 are transformed into three degree, and we verify whether
these transformed images are recognizable or not. Human subjects are asked to look at the transformed images
displayed on a tablet PC and to answer what they saw. Each character is nicked in several places in the image by a
white straight line. The width of the nick is gradually enlarged and we change the transformation degree of the
character image. The widths of the nicks are 10, 25, and 50 pixels in this experiment. We say that the transformation
degree is larger if the width of nick is bigger.
In addition, the character images are rotated clockwise 20 degree because recognition rate is expected to be
higher unless such transformation is applied. We consider that the characteristics of vertical and horizontal line are
not critical in feature demons. The transformed character images are shown in Fig. 2. The two digits following L, W,
s, and e indicate the width of nick. For example, the width of nick in L25 is 25 pixels. Note that these transformed
character images do not have the discontinuous curve feature of the Pandemonium architecture.

(a) L10 (b) W10 (c) s10 (d) e10

(e) L25 (f) W25 (g) s25 (h) e25

(i) L50 (j) W50 (k) s50 (l) e50

Fig. 2. Transformed character images

We show the subjects not only the transformed character images but also several dummy images as placeboes
because the subjects could forecast the original characters if they see the transformed character images in the order
 Mayumi Takaya et al. / Procedia - Social and Behavioral Sciences 97 (2013) 666 – 674 669

from the smaller transformation degree to the larger. The dummy images are chosen to represent some symbols or
characters so that the subjects cannot realize the purpose of the experiment. Twelve subjects, ten male
undergraduates and two female undergraduates, were asked to watch 12 images of transformed character images and
other 12 dummy images that have nothing to do with the characters L, W, s, and e. See Fig. 3 for the order of
images shown to the subjects in the experiment. A subject watches No.1 image W50 and then No. 2 and next No 3
dummy images, and so on. The subject is asked to watch images one by one until the subject watched No24 image.
After watching each image, each subject was asked to answer what he or she saw.
The experiment is carried out as follows.
(1) A subject is asked to watch the 24 images displayed on a tablet PC one by one and to answer what he or she
see. Each image is displayed at most five seconds. If the subject gives no answer, then a next image appears.
(2) After watching an image displayed on the tablet PC, the subject answers a recognized content, that is, what
the subject believes to have seen.
(3) The s recorded in each time.

No.1 W50 No.2 dummy No.3 dummy No. 4 s50 No.5 e50

No.6 dummy No.7 L50 No.8 dummy No.9 W25 No.10 dummy

No.11dummy No.12 s25 No.13 dummy No.14 e25 No.15 dummy

No.16 L25 No.17 w10 No.18 dummy No.19 dummy No.20 s10

No.21e10 No.22 dummy No.23 L10 No.24 dummy

Fig. 3. Order of images displayed

3.2. Experiment
E al result

We summarize the result in Table 2. It includes the name of images, the number of correct and wrong answers,
and the rate of correct answers. We ignore the answers to the dummy images, that is, No 2, 3, 6, 8, 10, 11, 13, 15, 18,
19, 22, and 24. All subjects answered correctly to the transformed character images with nicks of 10 pixels, that is,
L10, W10, s10 and e10. On the other hand, the rate of correct answers decreases for all transformed character
670 Mayumi Takaya et al. / Procedia - Social and Behavioral Sciences 97 (2013) 666 – 674

images with nicks of 50 pixels except for L50 has the 3 characteristics of the
Pandemonium architecture, that is, the vertical line, horizontal line, and right angle (see Table 1). We consider that
the rate is high for L50 because L50 still remains most of the characteristics of the vertical line, horizontal line, and
right angle even after inserting nicks.

Table 2. Experimental result

Image name L50 L25 L10 W50 W25 W10 s50 s25 s10 e50 e25 e10
Corresponding Character L L L W W W s s s e e e
Number of Correct Answer 12 12 12 0 11 12 4 12 12 8 12 12
Number of Wrong Answer 0 0 0 12 1 0 8 0 0 4 0 0
Correct Answer Rate (%) 100 100 100 0 91.7 100 33.3 100 100 66.7 100 100

No subject recognized the image W50 correctly. We consider that this happened since the characteristic of the
is completely lost by inserting nicks and therefore conclude that the acute angle is the most
important in the feature demons for The ratio of recognizing the image s50 is 33.3 percent. In the
image of s50, the curve disappears, and instead, 6 blocks appear by inserting nicks. Thus, we conclude that the
characteristic of the continuous curve . In summary, inserting nicks in the character images
affects feature demons of the Pandemonium architecture, although the influence depends on the characters. We
investigated only four characters in this experiment and so it is necessary to study in detail how inserting nicks
affects recognition ability by humans in future work.

4. Recognition by line trace simulators

4.1. Line trace simulator

We now examine the computer the shape and the order of a one-stroke character. Note
that we are interested in this study because of applications to the CAPTCHA. A line trace simulator is a computer
program behaving like a line trace robot and it may be able to trace a one-stroke character on a display. We use a
free software Line Trace Ver1.1 (downloaded from http://www.yamamoto-works.jp/). To analyze the recognition
ability, we try to trace transformed character images using the simulator. An image is given as a bitmap image, and
we set the color of the character the black (RGB values are all 0) and background the white (RGB values are all
255) as before. The size of the tracer is 20 10 pixel. The tracer follows pixels whose RGB values are all zero.
There are three virtual sensors in front of the tracer, and it moves straight or rotates itself right or left depending on
detected data by the sensors. Priority is given to moving straight. It retries to detect black pixels three times when it
fails to find a black pixel, and it stops unless it detects new black pixels. It can be adjusted the tracer for tracing
characters in our experiment but our aim is to clarify the primary difference between humans and computers and so
we did not tune up the simulator.

4.2. Experimental method

We use the images in Fig. 1 as basic images. Transformation is similar to the ones in Section 3.1 but no rotation
is applied. Since a line trace simulator is expected unable to trace a character with nicks if their width is relatively
large, we made images with nicks of finer width: 1, 3, 4, 10, 25, and 50 pixels, respectively. The line trace simulator
is executed on total 28 images: the basic images and 6 transformed images for each of 4 characters ,
and e ee Fig. 4). The two digits following L, W, s, and e indicate the width of nick, and 00 indicates there is no
nick as in Section 3.
We specify the position to start tracing for each transformed character before executing line trace simulation
because the line trace program does not know where it should start. The starting position is fixed for each character.
stroke order because the simulator has
multiple choices to go at the standard starting position and it would be bewildered for a line tracer. Line tracing
 Mayumi Takaya et al. / Procedia - Social and Behavioral Sciences 97 (2013) 666 – 674 671

succeeds only when the tracer passes all segments from the starting position to the terminal position in the correct
order and stops precisely at the terminal position. It ffails if the tracer stops on the way before the terminal position
or leaves away from the character segment. We accept only the correct one-stroke order. For example, see Fig. 5 for
the starting and terminal position of the character the blue rectangle indicates the tracer and it is placed on the
starting position when tracing begins, the green circle indicates the terminal position where the tracer is expected to
arrive eventually, and the red arrow indicates the correct order along which the tracer should follow. Note that our
motivation is to analyse the security of the CAPTCHA using one-stroke characters and so it is necessary to complete
tracing in the correct order in addition to recognizing characters.

L00 L01 L03 L04 L10 L25 L50

W00 W01 W03 W04 W10 W25 W50

s00 s01 s03 s04 s10 s25 s50

e00 e01 e03 e04 e10 e25 e50

Fig. 4. Character images for line trace simulator recognition

Tracer

Fig. 5. Tracing W00

4.3. Experimental result

We summarize the result of the line trace experiment in Table 3 where the obstacles for line tracer is indicated,
and we illustrate the start position by a red arrow and the position where the tracer stops by a blue rectangle for each
character images in Fig 6. It shows where the tracer stops in the case that it fails to trace. Tracing succeeds only in
the images s00 and s01. On the other hand, tracing fails in all the transformed character images of , ,
672 Mayumi Takaya et al. / Procedia - Social and Behavioral Sciences 97 (2013) 666 – 674

and the images s03, s04, s10, s25, and s50. In particular, tracing fails for transformed character images with nicks of
width larger than or equal to 3 pixels. Therefore we conclude that tracing can succeed only for character images
with nicks of width smaller than 3 pixels.

L00 L01 L03 L04 L10 L25 L50

W00 W01 W03 W04 W10 W25 W50

s00 s01 s03 s04 s10 s25 s50

e00 e01 e03 e04 e10 e25 e50

Fig. 6. Start position and the position when the tracer stops

We now discuss several cases that tracing fails. The tracer stops at the first nick of L10, W01, s10, and e04. In
addition, the tracer cannot turn a right angle at the corner of the characters L and e and stops. Similarly, the tracer
cannot turn on the acute angel of the character W, either. We should remark that we used the simulator with the
default setting. The line trace simulator can be tuned up for tracing characters to fill nicks and turn at right or acute
angles if we wish.
ed to recognize W50, s50 and
e50. On the other hand, humans recognized the other images L50, L25, L10, W25, W10, s25, s10, e25 and e10
although one subject failed to recognize W25. Humans were able to recognize all the transformed character images
tracer failed to recognize all
characteristic is tractable for humans to deal with but intractable for the line tracer. Likewise, humans recognized
line tracer did not in all of them. We
deduce the acute angle characteristic is the most influential feature of the char . If this characteristic is lost,
neither humans nor line tracer can retrieve it. Note that the image W50 lost completely the acute angle characteristic.
We should note that the tracer stops at the point where the character possesses the right or acute angle characteristic
of These characters do not have continuous curve feature and structured only by straight line.
T does not possess either acute or right angle feature. We should note that line tracing succeeded
only in s00 and s01 and it fails in all transformed character images possessing either acute or right angle
characteristic. Thus we conclude that the acute and right angle feature is one of obstacles for line tracers and the
continuous curve feature is tractable for line tracer.

and s10 is third, fourth and first nick, respectively. We do not know why this happened and so study more carefully
these cases in future work. To start the experiment of the character e , we placed the tracer at the point of the curve
 Mayumi Takaya et al. / Procedia - Social and Behavioral Sciences 97 (2013) 666 – 674 673

of lower right of the shape (see Fig. 6).

fact, the correct stroke order is to start at the junction. We guessed that it is hard for the tracer to recognize the
correct way to go. Therefore we placed the tracer at the end of the curve where the tracer can find only one way to
go.

Table 3. Experimental result

image L00 L01 L03 L04 L10 L25 L50

Trace Success - - - - - - -
Trace Failure
Obstacle right angle right angle 4th nick 4th nick 1st nick 1st nick 1st nick
image W00 W01 W03 W04 W10 W25 W50
Trace Success - - - - - - -
Trace Failure
Obstacle acute angle 1st nick 1st nick 1st nick 1st nick 1st nick 1st nick
image s00 s01 s03 s04 s10 s25 s50
Trace Success - - - - -

Trace Failure - -
Obstacle 3rd nick 4th nick 1st nick 1st nick 1st nick
image e00 e01 e03 e04 e10 e25 e50
Trace Success - - - - - - -
Trace Failure
Obstacle right angle right angle right angle 1st nick 1st nick 1st nick Lost control

4.4. Comparison with the existing study

We compare our experiment with the result by Tsuruta, Takaya and Yamamura [7] where humans and a line
trace simulator, which is the same as the one we used in this paper, try to respond to a CAPTCHA. To study
plausibility of attacks using a line trace simulator, the authors experimented how a line trace simulator acts to
character images given in Fig. 7. A tracer is placed on the starting position in each experiment because it does not
know where it should start. For their application, humans are required to trace one-stroke character on the
touchscreen as well. On the other hand, humans only have to recognize the character in our experiment in Section 2.
Humans succeeded in tracing all of the images of Fig. 7, but line tracing succeeded only in the image 4. The line
tracer failed to trace the image 1 because it does not know which way to go when it meets the junction while it
succeeded in the image 4 because it prefers to go straight when it meets the junction even though both characters
ne tracer prefers to go straight and does not return
the way and failed. It is intractable for a line tracer to behave like humans. It is a predictable result because the line
tracer does not have any mechanism to recognize characters or information on stroke order. On the other hand,
humans seem to recognize characters first and then carry out tracing and succeeded. These experiment shows
tracing one-stroke characters is not an easy task for a line tracer. In their study, the weakness of line trace simulator
is found such as image 1 and 2. In addition to their result, we found other weakness of line trace programs as
recognizer of one-stroke characters. These two studies are based on different experiment but findings are similar.

           
Image 1 Image 2 Image 3 Image 4 Image 5 Image 6

Fig. 7. Experiment images in [7]

674 Mayumi Takaya et al. / Procedia - Social and Behavioral Sciences 97 (2013) 666 – 674

Table 4. Experimental results in (Tsuruta, Takaya and Yamamura 2013)

Image 1 Image 2 Image 3 Image 4 Image 5 Image 6

Failure Failure Failure Success Failure Failure

5. Conclusions

Inspired by the research on a CAPTCHA based on recognizing one-stroke characters, we study recognition
ability of humans and computers. In particular, we consider one-stroke characters and its stroke order. We paid

recognition ability by experiments that humans and a line tracer recognize distorted and nicked character images. In
our experiments, a one-stroke character is nicked and so it does not seem the original character. Humans can
recognize such distorted and nicked character images with high probability, on the other hand a line trace program
cannot. In our experiment we only checked whether a distorted and nicked character is recognized as the original
character or not. From the point of view of applying to a CAPTCHA, it is left as a future research to study how
much humans can interpolate the nicks inserted in the character images. It is also a future research whether we can
tune up a line trace program to recognize and trace one-stroke characters like humans. The research is important for
security evaluation for the CAPTCHA technology. We also note that we did not deal with color issue in this paper.
Our study is incomplete in the sense that we investigate only four characters out of 26 alphabet characters.
The experimental results (Table 2 an Table 3) indicate
of computers by comparing experimental results and humans use not only pattern matching but also contextual
recognition ability to trace one-stroke characters. We set up a hypothesis that the significance of the feature demon
of Pandemonium architecture depends on characters. For example, the acute angle demon is more important than the
other demons to recognize the character W. It also seems that the acute angle and right angle demons are intractable
and continuous curve demon is tractable to a line trace simulator.

References

[1] Ahn von L., Blum M., Hopper N., Langford J. (2003). CAPTCHA: Using Hard AI Problems for Security. Eurocrypt 2003, LNCS 2656:
294-311.
[2] Ahn von L., Blum M., Langford J. (2004). Telling Humans and Computers Apart Automatically. Communications of the ACM 47, 56-
60.
[3] Ahn, von, L., Maurer, B., McMillen, C., Abraham, D., Blum, M. (2008). reCAPTCHA: Human-Based Character Recognition via Web
Security Measures. Science, 321, 1465-1468.
[4] Klatzky R.L. (1980). Human Memory; Structures and Procesees (Second Ed.) W. H. Freeman and Company
[5] Selfridge O. (1959). G. Pandemonium: A Paradigm for Learning. Proceedings of the Symposium on Mechanisation of Thought
Processes, 511-529.
[6] Selfridge O. G. and Neissser U. (1960). Pattern Recognition by machine. Scientific American 203, 60-67.
[7] Tsuruta Y., Takaya M., Yamamura A. (2013). CAPTCHA Suitable for Smartphones. ICT-EurAsia 2013, LNCS 7804, 131-140.
[8] Turing A.M. (1950). Computing Machinery and Intelligence. Mind 59 (236), 433-460