Hindawi

Security and Communication Networks

Volume 2017, Article ID 6898617, 15 pages
https://doi.org/10.1155/2017/6898617

Review Article
A Survey on Breaking Technique of Text-Based CAPTCHA

Jun Chen,1,2 Xiangyang Luo,1 Yanqing Guo,3 Yi Zhang,1 and Daofu Gong1
1
State Key Laboratory of Mathematical Engineering and Advanced Computing, Zhengzhou 450002, China
2
Henan Institute of Science and Technology, Xinxiang 453003, China
3
Dalian University of Technology, Dalian 116024, China

Correspondence should be addressed to Xiangyang Luo; luoxy [email protected]

Received 25 September 2017; Accepted 27 November 2017; Published 24 December 2017

Academic Editor: Zhenxing Qian

Copyright © 2017 Jun Chen et al. This is an open access article distributed under the Creative Commons Attribution License, which
permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

The CAPTCHA has become an important issue in multimedia security. Aimed at a commonly used text-based CAPTCHA, this
paper outlines some typical methods and summarizes the technological progress in text-based CAPTCHA breaking. First, the
paper presents a comprehensive review of recent developments in the text-based CAPTCHA breaking field. Second, a framework
of text-based CAPTCHA breaking technique is proposed. And the framework mainly consists of preprocessing, segmentation,
combination, recognition, postprocessing, and other modules. Third, the research progress of the technique involved in each
module is introduced, and some typical methods of segmentation and recognition are compared and analyzed. Lastly, the paper
discusses some problems worth further research.

1. Introduction abundant day by day. In 2013, [3] introduced CAPTCHAs of

the time and attacks against them; the authors investigated
As a multimedia security mechanism, CAPTCHA (Com- the robustness and usability of CAPTCHAs at the time
pletely Automated Public Turing test to tell Computers and and discussed ideas to develop more robust and usable
Humans Apart [1]) also called Human Interactive Proofs CAPTCHAs. Five years later, it is necessary to reorganize the
(HIP [2]), can protect multimedia privacy. Now, it has been emerging literature sources. Based on the research of text-
successfully applied to Google, Yahoo, Microsoft, and other based CAPTCHA breaking technique, this paper will review
major websites. In order to verify security and reliability the relative research and prospect future trends.
of CAPTCHA, the breaking technology came into being. The remainder of this paper is organized as follows: Sec-
It involves image processing, pattern recognition, image tion 2 briefly introduces the text-based CAPTCHA. Section 3
understanding, artificial intelligence, computer vision, and provides an overview of the text-based CAPTCHA breaking
many other disciplines. The research on CAPTCHA break- technique. Sections 4–8 describe main steps in the overall
ing has great value in research and application. First of framework of the text-based CAPTCHA breaking technique.
all, CAPTCHA breaking can verify the security of exist- Section 9 points out some problems which can be further
ing CAPTCHAs, and it can promote the development of studied. Section 10 concludes up the full manuscript.
CAPTCHA design technique. Secondly, the CAPTCHA is
an integral part of artificial intelligence and an important 2. Overview on Text-Based CAPTCHA
prerequisite to actualize natural human-computer interac-
tion. Finally, the research of breaking CAPTCHA not only In September 2000, the Carnegie Mellon University (CMU)
constantly refreshes limits to Turing test, but also can be research team designed the first commercial CAPTCHAs-
applied in other fields such as digital paper-based media, Gimpy series text-based CAPTCHAs to resist malicious
speech recognition, and image labeling. advertisements scattered by illegal scripting programs in
In recent decades, with the continuous development the Yahoo chat room. At the same time, the research on
of CAPTCHA technology, relevant literature sources are CAPTCHA design and breaking also started. In 2002 and
2 Security and Communication Networks

Table 1: Typical types of text-based CAPTCHA and their features.

Type Example Source Features

Character independent, texture background, some
Discuz!
interference

Slashdot A large number of interference lines and noise points

Solid
CAPTCHA Gimpy Multiple strings, overlap, distortion

Google Unfixed length, distortion, adhesion

Double-string, unfixed length, uneven thickness,

Microsoft
tilting, adhesion

QQ Hollow, shadows, interference shapes

Hollow
CAPTCHA Sina Hollow, adhesion, interference lines

Hollow, virtual contours, distortion, adhesion,

Yandex
interference lines

Scihub Hollow, shadows, interference lines, noise points

Three-
dimensional Grids, protrusion, distortion, background and
Teabag
CAPTCHA character blending
Colorful, shadow, rotation, zoom
Parc
Special characters
Program
Multiple characters jumping
Animation generating
CAPTCHA
Hcaptcha Multilayer character images blinking transformation

2005, the international seminars on HIP have been held, and (1) A large enough character set. Only when a character
a large number of related research results were published. set is large enough, the total number of CAPTCHA strings is
In subsequent years, many research results were reported large enough to resist violent breaking.
in international conferences including CVPR, NIPS, CCS, (2) The characters with distortion, adhesion, and overlap.
and NDSS. Many internationally renowned universities and Using characters with distortion, adhesion, and overlap, the
research institutions have established research groups on breaking methods cannot easily segmented a CAPTCHA
CAPTCHA technology, such as CMU [1, 8–14], PARC [15– image into single characters.
19], UCB [16, 17, 20, 21], Microsoft [2, 22–27], Google [28– (3) The characters are different in size, width, angle,
location, and fonts. When comparing features of different
30], Bell Laboratory [31, 32], Yan et al. [4, 33–42], Xidian Uni-
characters, the various transformations may reduce recogni-
versity [41–47], and University of Science and Technology of
tion accuracy.
China [48, 49]. In addition, many websites offer CAPTCHA (4) The strings with unfixed length. In a CAPTCHA
services in public such as CAPTCHA [10], BotBlock [50], scheme, strings with unfixed length can increase breaking
JCAPTCHA [51], and HCaptcha [52]. And some research difficulty to a certain extent.
groups focus on CAPTCHA recognition, such as PWNtcha (5) Hollow characters and broken contours. Compared
[53], Captchacker [54], aiCaptcha [55], and Gery Mori [56]. with the solid characters, hollow character’s features are less,
The security of text-based CAPTCHA mainly depends and broken contours can effectively resist the filling attack.
on the visual interference effects [25], including rotation, (6) The color and shape of complex backgrounds are
twisting, adhesion, and overlap. The typical types of text- similar to those of characters. If the images meet these
based CAPTCHA and their features are shown in Table 1. conditions, the noise is difficult to remove. This may reduce
To resist machine recognition, the text-based recognition accuracy.
CAPTCHA’s security is often protected by a series of The above features effectively enhance text-based
technologies. From Table 1, we can sum up the following CAPTCHAs’ security and bring great challenges to the
main features of the text-based CAPTCHA. CAPTCHA breaking research at the same time.
Security and Communication Networks 3

Table 2: Comparison of typical methods based on segmentation for breaking nonadherent CAPTCHA.

Example Source Success rate Reference Breaking method Year

Segmentation: character gap
Gimpy-r 78% [57] 2004
Recognition: distortion evaluation
Segmentation: connected region
EZ-Gimpy 97.9% [58] 2004
Recognition: distortion evaluation
Segmentation: vertical projection
Captcha-service 100% [34] Recognition: statistical character 2007
pixels
Segmentation: connected region
Ego-share 92.2% [5] 2009
Recognition: SVM

Ge-Captcha 100% [59] CW-SSIM 2010

Note. SVM: support vector machine, CW-SSIM: complex wavelet based structural similarity.

3. Research Progress of Breaking with the advantage of deep learning, the breaking based
Text-Based CAPTCHA on nonsegmentation will bounce back. The success rates of
typical text-based CAPTCHA breaking methods based on
For all kinds of text-based CAPTCHA schemes, the breaking nonsegmentation are as shown in Table 4.
methods are also various. According to whether there is
segmentation or not, the existing breaking methods be
3.3. The Framework of Text-Based CAPTCHA Breaking Tech-
contained in two categories.
nique. With the improvement of text-based CAPTCHA
design, the breaking technique changes to meet it. The early
3.1. Text-Based CAPTCHA Breaking Methods Based on Seg- text-based CAPTCHA contains nonadherent characters. The
mentation. The text-based CAPTCHA breaking based on breaking technique is the traditional framework of “prepro-
segmentation has different processing methods for different cessing + segmentation + recognition.” In recent years, most
objects and results. When there is no adherent character, indi- of the text-based CAPTCHAs use CCT (Crowded Characters
vidual characters are obtained using vertical projection and Together). Therefore, various breaking frameworks come
connected component with good effect. As shown in Table 2, into being, for example, “preprocessing + recognition,” “pre-
the success rates of nonadherent character CAPTCHA range processing + recognition + postprocessing,” “preprocessing
from 78% to 100%. + segmentation + combination + recognition,” and “pre-
However, it had little success in adherent characters. processing + segmentation + combination + recognition +
Therefore, more complicated methods, such as different postprocessing.”
width, character features, and character contours, have been In this paper, the existing frameworks are integrated
proposed one after another. With more and more antiseg- into an overall framework of text-based CAPTCHA break-
mentation technologies in CAPTCHA field, obtaining indi- ing, as shown in Figure 1. The framework mainly consists
vidual characters is becoming harder and harder. Then the of preprocessing, segmentation, combination, recognition,
researchers proposed the segmentation methods for obtain- postprocessing, and other modules. The research progress of
ing character components by character structure, filters, and each module will be described in the following.
so forth. As can be seen from Table 3, the success rates of
CAPTCHA breaking are generally low, with only a few higher
than 80%. 4. Preprocessing Methods of Breaking
Text-Based CAPTCHA
3.2. Text-Based CAPTCHA Breaking Methods Based on Non-
segmentation. The text-based CAPTCHA breaking methods The CAPTCHA preprocessing is the first step of CAPTCHA
based on nonsegmentation can directly recognize prepro- image processing before segmentation and recognition. Its
cessed CAPTCHA images. The breaking method’s success main purpose is to highlight the information related to
rate relies on recognition technique. In early stage, different characters in a given image and to weaken or eliminate inter-
pattern matching algorithms such as shape context [20] fering information. The preprocessing of existing CAPTCHA
and similarity [57] are used for recognition. Later, with breaking methods mainly includes image binarization, image
the improvement of the success rates of individual char- thinning, denoising, and so on.
acter recognition, researchers focus on the character seg-
mentation technique. However, the text-based CAPTCHA 4.1. Image Binarization. Image binarization is to highlight
design uses antisegmentation technique, which can prevent interesting objects’ contour and to remove noises in back-
obtaining complete and individual characters. Nowadays ground. The key to binarization is to select an appropriate
4 Security and Communication Networks

Table 3: Comparison of typical methods based on segmentation for breaking adherent CAPTCHA.

Example Source Success rate Reference Breaking method Year

Google, Segmentation: width
4.89%–66.2% [2] 2004
Yahoo Recognition: CNN
Microsoft 61%
Segmentation: color filling and projection
Google 8.7% [4] 2008
Recognition: CNN
Yahoo 25.9%
Segmentation: change width
Hotmail 40% [5] Recognition: SVM 2009
Post-processing: DP search
MSN 18%
[6] Segmentation: projection and central 2010
Yahoo 45%
Segmentation: color filling
Megaupload 78% [36] Combination: nonredundancy 2010
Recognition: CNN
reCAPT-
33% Segmentation: character structure feature
CHA [38] 2011
46.75% Recognition: CNN
Google
Segmentation: projection and character
Yahoo 54.7% [44] feature 2012
Recognition: OCR
Segmentation: color filling
Combination: redundancy
Yahoo 36%–89% [41] 2013
Recognition: CNN
Postprocessing: DFS
5.56% Different width/location segmenting and
Microsoft [60] 2015
57.05% template matching
reCAPT- Segmentation: trichromatic code
40.4%–94.3% [61] 2015
CHA Recognition: SVM
Edge and fuzzy logic segmentation and
Yahoo 57.3%–76.7% [7] 2015
recognition
Segmentation: Log-Gabor filter
Combination: redundancy
Microsoft 5%–77.2% [42] 2016
Recognition: KNN
Postprocessing: DP search
Segmentation: different width
MSN 27.1%–53.2% [48] 2016
Recognition: BPNN
Note. CNN: convolutional neural network, DP: dynamic programming, OCR: optical character recognition, DFS: depth first search, KNN: 𝑘-nearest neighbor,
BPNN: back-propagation neural network.

threshold. When the threshold is applied to the whole image, 4.3. Image Denoising. In order to resist breaking, there
it is called the global threshold method; otherwise, it is are noises and interference lines in CAPTCHA images.
called the local threshold method. If the threshold is not In addition, some noises are generated during grayscale
fixed during processing, it is called variable threshold method and binarization. Therefore, we need to denoise CAPTCHA
or dynamic threshold method. The common thresholding image. The typical methods are as shown in Table 5. We
methods are Sauvola and Pietikainen’s method [65], Otsu’s should choose the effective denoising method according to
method [66], and so on. actual situation.

4.2. Image Thinning. Image thinning is to process the char- 5. Segmentation Methods of Breaking
acter’s contour as skeleton. It must not change the character’s Text-Based CAPTCHA
adhesion. Its purpose is to highlight image contour and
to simplify subsequent processing. The thinning algorithms The segmentation aims to get individual characters or charac-
contain two categories: noniterative algorithm and itera- ter components. There are the segmentation methods based
tive algorithm. The common thinning algorithms include on individual characters and the segmentation methods
Hilditch algorithm [67] and Zhang and Suen algorithm [68]. based on character components.
Security and Communication Networks 5

Table 4: Comparisons of typical methods based on nonsegmentation for breaking adherent CAPTCHA.

Example Source Success rate Reference Breaking method Year

EZ-Gimpy 92% Shape context matching
[20] 2003
Gimpy 33% algorithm

EZ-Gimpy 99% [57] Correlation algorithm 2004

Program
55% [62] RNN 2011
generation
Program
54.9% [63] 2D LSTM-RNN 2013
generation

reCAPTCHA 99.8% [30] DCNN 2013

reCAPTCHA 31.75% [64] HMM 2015

Note. RNN: recurrent neural network, 2D LSTM: 2-dimensional long short-term memory, DCNN: spatial displacement of the neutral network, HMM: Hidden
Markov model.

5.1. Segmentation Methods Based on Individual Characters. 5.1.2. Segmentation Methods Based on Connected Components.
The segmentation methods based on individual characters The segmentation methods based on connected components
segment a CAPTCHA image to individual characters. For effectively segment individual characters using different con-
individual characters, we can use segmentation methods nected components in an image. For slope and distortion
based on character projection and connected components. characters, this method is effective. However, it is limited by
For CCT characters, we can use segmentation methods based adherent characters.
on character width, connected feature, and character contour. Reference [4] tried to segment Microsoft MSN
CAPTCHA by combining connected components and
vertical projection, as shown in Figure 3. First, different
5.1.1. Segmentation Methods Based on Character Projection. connected components are marked with different colors.
The segmentation methods based on character projection And then the character blocks are generated according to
determine the optimal segmentation position by analyzing different colors. Finally, strings are segmented to individual
the number of pixels projected under different conditions. characters using the vertical projection feature, with a success
This method applies to recognizing CAPTCHA characters rate of more than 90%.
without adhesion or slight adhesion. However, its effect is not
obvious for the seriously adherent and distorted characters. 5.1.3. Segmentation Methods Based on Character Width. The
The typical methods include vertical projection segmen- segmentation methods based on character width are suitable
tation, horizontal projection segmentation, and guideline for CAPTCHA images which are not easily segmented to
projection segmentation. individual characters. [60] used different widths (the average
Using (1), [61] defines three-color bar code to segment width of 0.75 times, 1 time, 1.5 times, and 2 times) to
reCAPTCHA images: segment an image. Thus, each character corresponds to four
recognition results, from which to find an optimal segment
for 𝐻Σ (𝑥) = 0, as the final recognition result. In addition, [5] did not take
{Blue,
{
{
{ the average width as standard; they gave a set of character
Three-color Bar (𝑥) {White, for 𝐻Σ (𝑥) = 1, (1) segments between the minimum width and the maximum
{
{
{ width and then determined the optimal segmentation scheme
{Black, for 𝐻Σ (𝑥) > 1, using dynamic programming, as shown in Figure 4.
where 𝐻Σ (𝑥) represents the total of object pixels in the
5.1.4. Segmentation Methods Based on Character Feature.
𝑥th column. In three-color bar a column is colored in blue The segmentation method based on character features uses
if there is not any pixel that belongs to character in the the features of CAPTCHA string, including inside features
column (𝐻Σ (𝑥) = 0). If there is only one pixel in column and outside features. Reference [38] classifies characters
(𝐻Σ (𝑥) = 1), the column is encoded by white. Finally, the according to their own inside features, and each class contains
black corresponds to the column with more than one object the characters as shown in Table 6.
pixel (𝐻Σ (𝑥) > 1), as shown in Figure 2(a). After denoising, Reference [6] segments characters according to outside
the optimal segmentation line is determined in the middle of features among them. This paper proposes a new seg-
blue bar or white bar, as shown in Figure 2(b). mentation algorithm called middle-axis point separation
6 Security and Communication Networks

Input images

Preprocessing
(i) Binarization
(ii) Thinning
(iii) Denoising
(iv) . . .

Yes No
Nonsegmentation?

Segmentation
Based on single character Based on character components
(i) Based on character projection (i) Based on character Structure
(ii) Based on connected components (ii) Based on filter
(iii) Based on character width (iii) . . .
(iv) Based on character feature
(v) Based on character contour
(vi) . . .

Yes No
Single character?

Combination
(i) Based on redundancy
(ii) Based on nonredundancy

Recognition

Based on template matching Based on character feature Based on machine learning

(i) Based on global property (i) Based on character structural feature (i) Based on traditional methods
(ii) Based on local feature (ii) Based on character statistical feature (ii) Based on neural network
(iii) Based on deep learning

Yes No
Redundancy?

Postprocessing
(i) Based on selection
(ii) Based on rejection
(iii) . . .

Output final results

Figure 1: The framework of text-based CAPTCHA breaking technique.

(a) Original three-color bar (b) Denoised three-color bar

Figure 2: Three-color bar corresponding to CAPTCHA image.

 Table 5: Comparisons of common denoising methods.
Denoising method Typical algorithm Implementation Advantages Disadvantages
The gray value of pixel is replaced by the
The irrelevant details and gaps are
Security and Communication Networks

Average filter mean of its neighboring pixels gray The image is blurred.
removed.
Denoising method values.
based on filter in the The gray value of pixel is replaced by the
Remove effectively the salt and
spatial domain Median filter median of its neighboring pixels gray Not applied to the image with many dots, lines, and spires.
pepper noise, speckle noise.
values.
The minimum mean square error Remove effectively the Gaussian
Wiener filter Computation is complex.
criterion is used to adjust the filter effect. noises.
Gibbs Markov random field theory. Remove effectively noise points.
Denoising method
The straight line in the image is detected Not applied to irregular interference line.
based on Gibbs and Remove effectively interference
Hough transform by using the point line duality of image
Hough transform lines.
space and Hough parameter space.
Smooth contours, cut off narrow
Open operation First corrosion to expansion. The effect of denoising varies with operation mode and the
Denoising method lines, and eliminate fine.
size and shape of structural elements; the experiment needs
based on morphology Smooth contour and fill holes, gaps,
Close operation First expansion to corrosion. to be repeated; the adaptability is poor.
and fracture of contour line.
The recursive method is used to find the
Remove effectively the noise
Denoising method connected domain to deal with pixel
Connected interference, and the original details Need to analyze character’s properties; hard to determine
based on connected points, and then denoising based on gray
component of the characters are generally not distinguish features.
component features and morphological features of
lost.
connected domain.
Denoising method Find the best mapping of original image
Complex computation and it needs to adjust relative
based on wavelet Wavelet transform in the wavelet transform domain to Retain more image details.
parameters.
transform restore the original image.
7
8 Security and Communication Networks

6. Combination Methods of Breaking

Text-Based CAPTCHA
An individual character after segmentation can be recognized
directly. But character components need to be combined
Figure 3: Segmented CAPTCHA image in [4]. into an individual character to be recognized. According to
the number of generated candidate characters, combination
technologies can be divided into two categories: the combi-
Table 6: Character class table.
nation technique based on redundancy and the combination
Class Dot Circle Cross S V technique based on nonredundancy.
Characters i, j a, b, d, e, g, o, p, q t, f s, z v, w, y
6.1. Combination Methods Based on Redundancy. The num-
ber of candidate characters generated by combination tech-
for CAPTCHAs. The algorithm utilizes the central pixel nique based on redundancy is more than the number of
in background between two disconnected object pixels as real characters. In [42], each character fragment is labeled
segmentation points (see Figure 5). in order from top to bottom and left to right, and then the
components are combined on the idea of jigsaw puzzle to
generate candidate characters.
5.1.5. Segmentation Methods Based on Character Contour.
The segmentation method based on character contours is to
analyze geometric features of character contours, so as to 6.2. Combination Methods Based on Nonredundancy. The
determine the appropriate segmentation lines. Reference [7] number of candidate characters generated by combination
tried to connect connection edge points between two merged technique based on nonredundancy is equal to the number
characters and determined the optimal segmentation line by of actual characters. In [36], the character components are
confidence, as shown in Figure 6. nonredundant. The overlap area strokes may be reused to
compose a complete character. Figure 7(a) shows a Megau-
pload CAPTCHA image. Figure 10 gives the combined four
5.2. Segmentation Methods Based on Character Components. characters. The final success rate of combination is 78.25%.
The segmentation methods based on character components
produce multiple character components, rather than individ-
ual characters. The segmentation methods are mainly base on 7. Recognition Methods of Breaking
character structure or filter. Text-Based CAPTCHA
Nowadays, the recognition methods used in text-based
5.2.1. Segmentation Methods Based on Character Structure. CAPTCHA system include three categories: template match-
Using structural feature of characters with black compo- ing, character feature, and machine learning.
nents and white components, [36] segmented a seriously
overlapped string to multiple components. First, locate black
components, as shown in Figure 7(b). And then, locate 7.1. Recognition Methods Based on Template Matching. Tem-
white components, as shown in Figure 7(c). Finally, identify plate matching is to compare similarity of each pixel between
black components of each character and the shared white characters and every template and to find the highest simi-
components. larity. According to matching range, there are the matching
In [41], a CAPTCHA image contains several hollow recognition methods based on global property and the
characters, whose contours naturally form several closed matching recognition methods based on local feature.
regions (see Figure 8(a)). According to this structural feature,
a character is segmented to several character components by 7.1.1. Matching Recognition Methods Based on Global Property.
color filling (see Figure 8(b)). The matching recognition methods based on global property
is traverse scanning. Within search area, the optimal match
5.2.2. Segmentation Methods Based on Filter. Reference [42] point to each pixel is found by regional correlation matching
is the first to apply Gabor filters for breaking CAPTCHAs, calculation. Because many templates matching each pixel will
which extracts character components along four directions be pretty slow, [45] proposes the second template matching
by convolving a CAPTCHA image with each of four filters, algorithm to improve efficiency. Only if a rough matching is
respectively, as shown in Figure 9. The segmentation method successful, an exact matching needs to be made.
is not limited by adhesion, distortion, and overlap and is
suitable for many kinds of characters. 7.1.2. Matching Recognition Methods Based on Local Feature.
In summary, the contrast among segmentation methods The shape context is a simple local feature shape descriptor.
is given. As can be seen in Table 7, each segmentation Its basic idea is to convert the matching problem of image
method applies to different types of characters. It is only the into the matching problem of feature point set. In 2003, Mori
individualized segmentation method that can obtain good and Malik [20] used shape context to break the CAPTCHA
results. of Gimpy and EZ-Gimpy. For good robustness to image
Security and Communication Networks 9

３3 ３4 ３13 ３14 ３19

３2 ３5 ３12 ３15
３6 ３11 ３16 ３20
３1 ３21
３10 ３17
３7 ３9 ３18 ３22
３8 ３23

Figure 4: Character segments in [5].

Table 7: Comparisons of segmentation methods.

Character features
Segmentation methods Main basis Explanation
Adhesion Skew Distortion Overlap
Suit strings with little adhesion rather
Character projection √ × × × than serious distorted, overlapped
strings
Connected Limited to overlapping rather than
× √ √ ×
Segmentation based on components distortion
individual character Character width √ × × × Limited to severely distorted
Effective when character feature is
Character feature √ × × ×
notable and discriminant
Applied to the individual character
Character contour √ √ √ ×
with clear contour
Only when character structure is easy
Character structure √ √ √ √
Segmentation base on to segment
character components Wide application; the subsequent
Filter √ √ √ √
processing is complex

7.2. Recognition Methods Based on Character Feature.

Because the character of each CAPTCHA mechanism varies
in design, we can define different methods according to the
feature of characters, which is mainly based on character
Figure 5: The middle-axis points in [6]. structural feature and character statistical feature.

QuadI QuadII 7.2.1. Recognition Methods Based on Character Structural

Feature. The structural feature can describe the details and
structural information of characters, such as the number of
loops, inflection point, convexo-concave degree, and cross
points. For example, [46] uses the guidelines of characters
(see Figure 11(a)) and closed loop detection (see Figure 11(b))
to break Yahoo CAPTCHA.

7.2.2. Recognition Methods Based on Character Statistical Fea-

ture. The recognition method based on character statistical
feature uses commonly statistical features including pixel
feature, projection feature, contour feature, and coarse mesh
feature. This feature is robust to noise interference and is
widely used in CAPTCHA recognition field. Reference [34]
used the distinct pixel count for each of the letters A to Z (see
QuadIII QuadIV Figure 12) to break captchaservice.org CAPTCHA with a near
Figure 6: The optimal segmentation line in [7]. 100% success rate.

7.3. Recognition Methods Based on Machine Learning. The

scaling and affine transformation, it is widely used in face recognition methods based on machine learning is essen-
recognition, CAPTCHA recognition, shape matching, and tially using machine learning algorithms to correctly classify
other fields. CAPTCHA characters. According to chronological order of
10 Security and Communication Networks

(a) Original image (b) Nonshared character compo- (c) Shared character components
nents
Figure 7: An example of segmented CAPTCHA image in [36].

(a) Original image (b) Segmented image

Figure 8: An example of segmented CAPTCHA image in [41].

Filter Binarize
(0)

(/4)

(/2)

(3/4)

Gabor
filters

Figure 9: Segmentation CAPTCHA image by Gabor filters in [42].

achieved higher success rates on most of the schemes, but

CNN was faster most of the time.

7.3.2. Recognition Methods Based on Neural Network. For the

principle of parallel distributed operation in large number of
Figure 10: Individual characters after combination in [36]. neurons, the efficient learning algorithms, and the ability to
imitate human cognitive systems, the neural network is very
suitable to solve problems such as speech recognition and text
mainstream, it can be basically divided into three categories: recognition.
traditional methods, neural network, and deep learning. In [62], a BPNN used cross entropy for calculating
the performance of a network with targets and outputs.
7.3.1. Recognition Methods Based on Traditional Methods. In Eventually, the system achieved an overall precision of 51.3%,
the field of text-based CAPTCHA recognition, the most 27.1%, and 53.2% for the CCT CAPTCHAs of Taobao, MSN,
widely used traditional classifiers include SVM and KNN. and eBay, respectively.
The idea of SVM is to separate classes via a hyperplane. However, when applying neural network, we need to
The key is kernel function, which is responsible for mapping extract character features first. The quality of extracted
original features into high-dimensional space in a nonlinear features limits the final recognition rate to a certain extent.
way, thereby improving the separability for data. Reference
[5] compared four kernel functions: RBF (Radius-Based 7.3.3. Recognition Methods Based on Deep Learning. In recent
Function), POLY (polynomial), LINEAR, and SIGMOID. years, deep learning has achieved remarkable achievements
The experimental results showed that the performance of the in recognition fields of text, image, audio, and so forth.
first two kernel functions was optimal. The deep learning models commonly used in CAPTCHA
KNN is based on the category of the nearest 𝐾 samples recognition field are CNN, RNN, LSTM-RNN, and so forth.
to determine the category of a sample. Reference [42] tested CNN recognizes character images without feature extrac-
SVM, BPNN (back-propagation neural network), template tion and has a certain degree of robustness in displacement,
matching, CNN, and KNN. Among these classifiers, KNN scale, and deformation. In the existing research results, a
Security and Communication Networks 11

(a) The guideline (b) The closed loop detection

Figure 11: The example of character structural features in [46].

300 8.1. Postprocessing Methods Based on Selection. Usually, there

are many redundant individual characters generated in pre-
250 M
N W vious steps. This requires selecting the most likely combined
B Q
R
string as the final recognition result of CAPTCHA image.
200 A D GH S Z The selection strategies include the local optimization and the
O X
K TU global optimization.
Pixel count

C E P V Y The local optimization selection only takes into account

150 F
I J
the recognition confidence optimality of an individual char-
L acter. In [60], each character corresponds to several candidate
100 characters with different widths. Therefore, the candidate
character with the highest recognition confidence is selected
50 as the final character.
The global optimization selection strives for the best
0 results for all characters in an image. In [41], all candi-
Letter date characters are found by the graph traversal, and then
Figure 12: The pixel count for each of the letters A to Z in [34]. the string with the highest sum of characters recognition
confidence values is taken as the final result, while in
[5, 42], to avoid enumerating all candidate characters, a
dynamic programming is used to determine the final result
typical CNN is widely used [2, 4, 36, 38, 41] with a good recog- with the highest sum of characters’ recognition confidence
nition accuracy. Reference [30] trained large, distributed deep values directly. Compared with graph traversal, the dynamic
convolutional neural networks and achieved 99.8% accuracy programming is more effective and accurate.
in recognizing CAPTCHA images of reCAPTCHA.
However, due to lack of time dimension, CNN cannot 8.2. Postprocessing Methods Based on Rejection. The purpose
combine context information in recognition. So RNN with of postprocessing methods based on rejection is to determine
feedback and time parameters was proposed to process time whether the tested sample belongs to the types of training
series data. Later, in order to solve vanishing gradient prob- set by analyzing character recognition results. Therefore, the
lem of RNN, LSTM was proposed in machine learning field. postprocessing methods based on rejection are a key to
Reference [62] applied 2D LSTM-RNN in CCT CAPTCHAs ensure high reliability of CAPTCHA recognition.
recognition with a success rate of 55.2%. It innovatively At present, the researchers have not been paid enough
obtained relative information not only in horizontal context, attention to the postprocessing methods based on rejection.
but also in vertical context. To the best of our knowledge, there is only one paper [62]
In summary, a contrast among recognition methods is in CAPTCHA field. It considers multiple features, such as
given, as shown in Table 8. According to the features of confidence, string length, character spaces, and the first
different networks, we should attempt to construct a new and the last character of a string, to determine whether a
deep learning model by combining multiple networks. It will candidate character should be rejected or not.
be a research trend in the field of text-based CAPTCHA
recognition.
9. Some Problems Worth Further Research
8. Postprocessing Methods of Breaking As stated above, many achievements have been acquired.
Text-Based CAPTCHA However, in view of the complexity of text-based CAPTCHA,
there are still some issues worth exploring in depth in this
In previous steps, some of character recognition results field.
may be taken as final results directly, while others need to
be further postprocessed. In postprocessing stage, the final (1) Construction of Standard Test Database for Text-Based
result’s reliability is ensured by simplification, selection, and CAPTCHA. A rich and high quality text-based CAPTCHA
optimization. According to different objects and methods, image database is the necessary foundation for the research of
there are the postprocessing methods based on selection and text-based CAPTCHA breaking. At present, the researchers
the postprocessing methods based on rejection. get CAPTCHA images mainly by web access and software
 12

Table 8: Comparisons of recognition methods.

Recognition methods Main basis Typical methods Advantages Disadvantages
The program is simple and suitable for The required template library is large; it
Recognition method Global property Traversal search matching algorithm
standard character verification code. depends on the choice of template matching.
based on template
The image information is rich, and it is robust
matching Local feature Shape context matching algorithm Without rotation invariant.
to image scaling and affine transformation.
Character structure Algorithm based on character structure Sensitive to the details of characters; strong in The distortion is serious when there are noise
Recognition method
feature feature distinguishing features. interferences.
based on character
Character statistical Algorithm based on character statistical
feature Strong robustness against noise interference. Targeted; application limited.
feature feature
Strong approximation ability and
generalization ability; good adaptability and
SVM Not applied to infinite sample space.
high accuracy for small sample space; suitable
Template matching for two kinds of classification.
It is better to avoid the problem of imbalanced
Computation is complex; easy to misjudge in
KNN samples, which is suitable for overlapping
Recognition method the domain with small sample size.
samples of the same class.
based on machine Flexible structure design, suitable for multiclass
learning Traditional method BPNN Slow convergence rate; depends on parameters.
classification.
Accepts an input image directly; automatically
extracts features; own robustness to Lack of time dimension; it could not identify
CNN
displacement, scale, and deformation; high using context information.
Deep learning recognition accuracy.
RNN Processes data in time series. Time gradient may disappear.
Owns the time memory function; effective to
LSTM-RNN Unable to extract feature automatically.
prevent gradient disappear.
Security and Communication Networks
Security and Communication Networks 13

generation. However, due to the diversity and timeliness of technique. First of all, this paper introduces various text-
text-based CAPTCHA, it has not been possible to construct a based CAPTCHAs and focuses on their features. Second,
common image database in the field of text-based CAPTCHA according to whether there is segmentation or not, we classify
recognition. It is necessary to collect, classify, organize, and the existing breaking methods of text-based CAPTCHA
establish the text-based CAPTCHA images database. The and summarize their features. Meanwhile, we propose a
database can provide the reliable training and testing data framework of text-based CAPTCHA breaking technique and
for research work and also provide the premise and basis of introduce the modules contained in the framework one
unified evaluation for various methods in this field. by one. Next, we compare and analyze the basic princi-
ples, advantages, and disadvantages of the existing methods
(2) Multitype CAPTCHA Recognition. At present, only when from five aspects: preprocessing, segmentation, combination,
training set and test set belong to the same type, the classifier recognition, and postprocessing. Finally, some problems
can effectively recognize CAPTCHAs. In fact, there are a worth further research are discussed.
variety of character changes in a CAPTCHA. Therefore, it
is an arduous and important task to design a reasonable
classifier to recognize various types of CAPTCHAs.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
(3) Segmentation-Free CAPTCHA Recognition. After more
than ten years of development, the text-based CAPTCHA
breaking has achieved a high success rate in individual char- Acknowledgments
acter. However, the breaking success rate of the CAPTCHA
This work was supported by the National Natural Science
string is generally low, and the results are less. With the
Foundation of China (nos. 61379151, 61401512, 61572052,
wide application of CCT strings in text-based CAPTCHA, the
and U1636219), the National Key R&D Program of China
problem of segmentation-free CAPTCHA recognition needs
(nos. 2016YFB0801303 and 2016QY01W0105), and the
to be solved urgently. Now deep learning may provide new
Key Technologies R&D Program of Henan Province (no.
ideas and technical means to solve this problem.
162102210032).
(4) Application of Deep Learning Model. At present, in the
field of CAPTCHA recognition, deep learning model can References
achieve better results than traditional methods. The most
frequently used methods are based on CNN and its improved [1] L. Von Ahn, M. Blum, and J. Langford, “Telling humans and
methods, while other deep learning models such as DBN computers apart automatically,” Communications of the ACM,
vol. 47, no. 2, pp. 56–60, 2004.
(Deep Belief Networks), RNN, LSTM/BLSTM/MDLSTM,
[2] K. Chellapilla and P. Y. Simard, “Using Machine Learning to
and DRL (Deep Reinforcement Learning) were not well
Break Visual Human Interaction Proofs (HIPs),” in Proceedings
used in text-based CAPTCHA recognition. Furthermore, of the Advances in Neural Information Processing Systems,
the study of the interrelationships and fusion applications pp. 265–272, ofAdvances in Neural Information Processing
between the various deep learning models is not thorough. Systems, 2004.
We hope that newer and better deep learning models are [3] N. Roshanbin and J. Miller, “A survey and analysis of current
proposed to make a breakthrough in CAPTCHA recognition, CAPTCHA approaches,” Journal of Web Engineering, vol. 12, no.
which will certainly promote the development in this field. 1-2, pp. 001–040, 2013.
[4] J. Yan and A. S. E. Ahmad, “A low-cost attack on a microsoft
(5) Rejection of Text-Based CAPTCHA. With the development CAPTCHA,” in Proceedings of the 15th ACM conference on
of CAPTCHA breaking technique, the reliability of recogni- Computer and Communications Security, CCS’08, pp. 543–554,
tion results is also increasing. In this regard, on one hand, we USA, October 2008.
should improve the correct rate of recognition; on the other [5] F. Jean-Baptiste and R. Paucher, “The Captchacker Project,”
hand, we should guarantee the correct rejection. In the field 2009, http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1
of CAPTCHA recognition, the concept of rejection has not .1.800.3065&rep=rep1&type=pdf.
been well known to the researchers. Therefore, this study has [6] S.-Y. Huang, Y.-K. Lee, G. Bell, and Z.-H. Ou, “An efficient
a potential development space. segmentation algorithm for CAPTCHAs with line cluttering
and character warping,” Multimedia Tools and Applications, vol.
(6) Misrecognition of Confusable Characters. When using the 48, no. 2, pp. 267–289, 2010.
deep learning network to extract character features automat- [7] R. A. Nachar, E. Inaty, P. J. Bonnin, and Y. Alayli, “Breaking
ically, the characters with similar features are easily confused. down Captcha using edge corners and fuzzy logic segmen-
It has practical significance to improve the precision of feature tation/recognition technique,” Security and Communication
extraction and the training methods in the deep learning Networks, vol. 8, no. 18, pp. 3995–4012, 2015.
network. [8] L. von Ahn, M. Blum, N. J. Hopper, and J. Langford,
“CAPTCHA: using hard AI problems for security,” in Advances
in cryptology—EUROCRYPT 2003, vol. 2656 of Lecture Notes in
10. Conclusions Computer Science, pp. 294–311, Springer, Berlin, Germany, 2003.
Based on detailed investigation and in-depth analysis, this [9] https://www.google.com/recaptcha.
paper reviews the progress of text-based CAPTCHA breaking [10] http://captcha.net/.
14 Security and Communication Networks

[11] http://www.captcha.net/captchas/bongo. [27] Y. Rui and Z. Liu, “ARTiFACIAL: Automated reverse turing test
[12] A. Schlaikjer and A. Dual, “Use Speech CAPTCHA: Aiding using FACIAL features,” Multimedia Systems, vol. 9, no. 6, pp.
Visually Impaired Web Users while Providing Transcriptions of 493–502, 2004.
Audio Streams,” Tech. Rep. LTI-CMU-07-014, Carnegie Mellon [28] K. A. Kluever and R. Zanibbi, “Balancing usability and security
University, Pittsburgh, Pa, USA, 2007. in a video CAPTCHA,” in Proceedings of the 5th Symposium On
[13] J. Tam, J. Simsa et al., “Improving Audio CAPTCHAs,” in Usable Privacy and Security, SOUPS 2009, USA, July 2009.
Proceedings of the Symposium on Usable Privacy and Security, [29] R. Gossweiler, M. Kamvar, and S. Baluja, “What’s up
2008. CAPTCHA? A CAPTCHA based on image orientation,”
[14] J. Tam, S. Hyde, J. Simsa, and L. Von Ahn, “Breaking audio in Proceedings of the 18th International World Wide Web
CAPTCHAs,” in Proceedings of the 22nd Annual Conference on Conference, WWW 2009, pp. 841–850, Spain, April 2009.
Neural Information Processing Systems, NIPS 2008, pp. 1625– [30] I. J. Goodfellow, Y. Bulatov, J. Ibarz et al., “Multi-digit
1632, can, December 2008. Number Recognition from Street View Imagery using Deep
[15] H. S. Baird and K. Popat, “Human Interactive Proofs and Convolutional Neural Networks,” 2014, https://www.research-
Document Image Analysis,” in Proceedings of the International gate.net/publication/259399973 Multi-digit Number Recog-
Workshop on Document Analysis Systems, vol. 2423 of Lecture nition from Street View Imagery using Deep Convolutional
Notes in Computer Science, pp. 507–518, Springer, 2002. Neural Networks.
[16] A. L. Coates, H. S. Baird, and R. J. Fateman, “Pessimal print: [31] T.-Y. Chan, “Using a test-to-speech synthesizer to generate a
A reverse turing test,” in Proceedings of the 6th International reverse Turing test,” in Proceedings of the 15th IEEE International
Conference on Document Analysis and Recognition, ICDAR Conference on Tools with Artificial Intelligence, pp. 226–232,
2001, pp. 1154–1158, usa, September 2001. Sacramento, Calif, USA, 2003.
[17] M. Chew and H. S. Baird, “Baffletext: A human interactive [32] G. Kochanski, D. Lopresti, and C. Shih, “A reverse turing test
proof,” in Proceedings of the Document Recognition and Retrieval using speech,” in Proceedings of the 7th International Conference
X, pp. 305–316, USA, January 2003. on Spoken Language Processing, ICSLP 2002, pp. 1357–1360,
[18] R. Chow, P. Golle, M. Jakobsson, L. Wang, and X. Wang, September 2002.
“Making CAPTCHAs clickable,” in Proceedings of the 9th Work- [33] http://www.lancaster.ac.uk/people/yanj2/.
shop on Mobile Computing Systems and Applications, HotMobile [34] J. Yan and A. S. El Ahmad, “Breaking visual CAPTCHAs
2008, pp. 91–94, USA, February 2008. with naı̈ve pattern recognition algorithms,” in Proceedings of
[19] P. Golle, “Machine learning attacks against the asirra the 23rd Annual Computer Security Applications Conference,
CAPTCHA,” in Proceedings of the 15th ACM conference on ACSAC 2007, pp. 279–291, December 2007.
Computer and Communications Security, CCS’08, pp. 535–542, [35] J. Yan and A. S. El Ahmad, “Usability of CAPTCHAs or
USA, October 2008. usability issues in CAPTCHA design,” in Proceedings of the 4th
[20] G. Mori and J. Malik, “Recognizing objects in adversarial Symposium on Usable Privacy and Security, SOUPS 2008, pp.
clutter: breaking a visual CAPTCHA,” in Proceedings of the IEEE 44–55, July 2008.
Computer Society Conference on Computer Vision and Pattern [36] A. S. El Ahmad, J. Yan, and L. Marshall, “The robustness of a
Recognition, vol. 1, pp. 134–144, June 2003. new CAPTCHA,” in Proceedings of the 3rd European Workshop
[21] M. Chew and J. D. Tygar, “Image Recognition CAPTCHAs,” on System Security, EUROSEC’10, pp. 36–41, April 2010.
in Proceedings of the 7th International Information Security
[37] B. B. Zhu, J. Yan, Q. Li et al., “Attacks and design of image
Conference, vol. 3225 of Lecture Notes in Computer Science, pp.
recognition CAPTCHAs,” in Proceedings of the 17th ACM
268–279, Springer.
Conference on Computer and Communications Security, CCS’10,
[22] K. Chellapilla, K. Larson, P. Simard, and M. Czerwinski, pp. 187–200, October 2010.
“Designing human friendly human interaction proofs (HIPs),”
[38] A. S. E. Ahmad, J. Yan, and M. Tayara, “The Robustness of
in Proceedings of the the SIGCHI conference, p. 711, Portland,
Google CAPTCHAs,” Computing Science Technical Report CS-
Oregon, USA, April 2005.
TR-1278, Newcastle University, 2011.
[23] P. Y. Simard, R. Szeliski, J. Benaloh, J. Couvreur, and I. Calinov,
[39] A. S. El Ahmad, J. Yan, and W.-Y. Ng, “CAPTCHA design: Color,
“Using character recognition and segmentation to tell computer
usability, and security,” IEEE Internet Computing, vol. 16, no. 2,
from humans,” in Proceedings of the 7th International Conference
pp. 44–51, 2012.
on Document Analysis and Recognition, ICDAR 2003, pp. 418–
423, UK, August 2003. [40] A. Algwil, D. Ciresan, B. Liu, and J. Yan, “A security analysis
of automated Chinese turing tests,” in Proceedings of the 32nd
[24] K. Chellapilla, K. Larson, P. Y. Simard, and M. Czerwinski,
Annual Computer Security Applications Conference, ACSAC
“Building segmentation based human-friendly human interac-
2016, pp. 520–532, December 2016.
tion proofs (HIPs),” in Proceedings of the Second International
Workshop on Human Interactive Proofs, HIP 2005, pp. 1–26, usa, [41] H. Gao, W. Wang, J. Qi, X. Wang, X. Liu, and J. Yan, “The
May 2005. robustness of hollow CAPTCHAs,” in Proceedings of the ACM
[25] K. Chellapilla, K. Larson, P. Simard, and M. Czerwinski, SIGSAC Conference on Computer and Communications Security,
“Computers beat humans at single character recognition in CCS 2013, pp. 1075–1085, November 2013.
reading based human interaction proofs (HIPs),” in Proceedings [42] H. Gao, J. Yan, F. Cao et al., “A Simple Generic Attack on Text
of the 2nd Conference on Email and Anti-Spam, usa, July 2005. Captchas,” in Proceedings of the Network and Distributed System
[26] J. Elson, J. R. Douceur, J. Howell, and J. Saul, “Asirra: A Security Symposium, pp. 1–14, San Diego, Calif, USA, 2016.
CAPTCHA that exploits interest-aligned manual image cat- [43] http://web.xidian.edu.cn/hchgao/paper.html.
egorization,” in Proceedings of the 14th ACM Conference on [44] H. Gao, W. Wang, and Y. Fan, “Divide and conquer: An
Computer and Communications Security, CCS’07, pp. 366–374, efficient attack on Yahoo! CAPTCHA,” in Proceedings of the 11th
USA, November 2007. IEEE International Conference on Trust, Security and Privacy
Security and Communication Networks 15

in Computing and Communications, TrustCom-2012, pp. 9–16, audio symbols,” Journal of Information Processing, vol. 23, no. 6,
June 2012. pp. 814–826, 2015.
[45] F. Dai, H. Gao, and D. Liu, “Breaking CAPTCHAs with second [65] J. Sauvola and M. Pietikäinen, “Adaptive document image
template matching and BP neural network algorithms,” Interna- binarization,” Pattern Recognition, vol. 33, no. 2, pp. 225–236,
tional Journal of Information Processing and Management, vol. 4, 2000.
no. 3, pp. 126–133, 2013. [66] N. Otsu, “A threshold selection method from gray-level his-
[46] H. Gao, W. Wang, Y. Fan, J. Qi, and X. Liu, “The robustness tograms,” IEEE Transactions on Systems, Man, and Cybernetics,
of “connecting characters together” CAPTCHAs,” Journal of vol. 9, no. 1, pp. 62–66, 1979.
Information Science and Engineering, vol. 30, no. 2, pp. 347–369, [67] C. J. Hilditch, “Linear Skeletons from Square Cupboards,”
2014. Machine Intelligence, pp. 403–420, 1969.
[47] H. Gao, X. Wang, F. Cao et al., “Robustness of text-based [68] T. Y. Zhang and C. Y. Suen, “A fast parallel algorithm for
completely automated public turing test to tell computers and thinning digital patterns,” Communications of the ACM, vol. 27,
humans apart,” IET Information Security, vol. 10, no. 1, pp. 45– no. 3, pp. 236–239, 1984.
52, 2016.
[48] R. Hussain, H. Gao, and R. A. Shaikh, “Segmentation of
connected characters in text-based CAPTCHAs for intelligent
character recognition,” Multimedia Tools and Applications, pp.
1–15, 2016.
[49] R. Hussain, H. Gao, R. A. Shaikh, and S. P. Soomro, “Recog-
nition based segmentation of connected characters in text
based CAPTCHAs,” in Proceedings of the 8th IEEE International
Conference on Communication Software and Networks, ICCSN
2016, pp. 673–676, June 2016.
[50] https://captcha.com/.
[51] http://jcaptcha.sourceforge.net/.
[52] http://www.hinsite.com.
[53] http://caca.zoy.org/wiki/PWNtcha.
[54] https://code.google.com/p/captchacker.
[55] http://www.brains-n-brawn.com/default.aspx?vDir=aicaptcha.
[56] http://www.cs.sfu.ca/∼mori/research/gimpy/.
[57] G. Moy, N. Jones, C. Harkless, and R. Potter, “Distortion estima-
tion techniques in solving visual CAPTCHAs,” in Proceedings of
the IEEE Computer Society Conference on Computer Vision and
Pattern Recognition, CVPR 2004, pp. II23–II28, July 2004.
[58] A. Bansal, D. Garg, and A. Gupta, “Breaking a Visual
CAPTCHA: A Novel Approach using HMM,” 2008, https://pdfs
.semanticscholar.org/3c2c/9af1e9a3b7095edaf8f205dfbadc30f-
917fb.pdf.
[59] S. Li, S. A. H. Shah, M. A. U. Khan, S. A. Khayam, A.-R.
Sadeghi, and R. Schmitz, “Breaking e-banking CAPTCHAs,” in
Proceedings of the 26th Annual Computer Security Applications
Conference, ACSAC 2010, pp. 171–180, December 2010.
[60] C. Hong, B. Lopez-Pineda, K. Rajendran, and A. Recasens,
“Breaking Microsoft’s CAPTCHA,” 2015, https://courses.csail
.mit.edu/6.857/2016/files/hong-lopezpineda-rajendran-recan-
sens.pdf.
[61] O. Starostenko, C. Cruz-Perez, F. Uceda-Ponga, and V. Alarcon-
Aquino, “Breaking text-based CAPTCHAs with variable word
and character orientation,” Pattern Recognition, vol. 48, no. 4,
pp. 1097–1108, 2015.
[62] L. Zhang, L. Zhang, S.-G. Huang, and Z.-X. Shi, “A highly
reliable CAPTCHA recognition algorithm based on rejection,”
Acta Automatica Sinica, vol. 37, no. 7, pp. 891–900, 2011.
[63] R. Chen, J. Yang, R.-G. Hu, and S.-G. Huang, “A novel LSTM-
RNN decoding algorithm in CAPTCHA recognition,” in Pro-
ceedings of the 3rd International Conference on Instrumentation
and Measurement, Computer, Communication and Control,
IMCCC 2013, pp. 766–771, September 2013.
[64] S. Sano, T. Otsuka, K. Itoyama, and H. G. Okuno, “HMM-based
attacks on Google’s ReCAPTCHA with continuous visual and
 International Journal of

Rotating
Machinery

International Journal of
The Scientific
(QJLQHHULQJ Distributed
Journal of
Journal of

Hindawi Publishing Corporation

World Journal
Hindawi Publishing Corporation Hindawi Publishing Corporation
Sensors
Hindawi Publishing Corporation
Sensor Networks
Hindawi Publishing Corporation
http://www.hindawi.com Volume 201 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Journal of

Control Science
and Engineering

Advances in
Civil Engineering
Hindawi Publishing Corporation Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Submit your manuscripts at

https://www.hindawi.com

Journal of
Journal of Electrical and Computer
Robotics
Hindawi Publishing Corporation
Engineering
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

VLSI Design
Advances in
OptoElectronics
,QWHUQDWLRQDO-RXUQDORI

International Journal of
Modelling &
Simulation
$HURVSDFH
Hindawi Publishing Corporation Volume 2014
Navigation and
Observation
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014
in Engineering
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014
(QJLQHHULQJ
+LQGDZL3XEOLVKLQJ&RUSRUDWLRQ
KWWSZZZKLQGDZLFRP 9ROXPH
Hindawi Publishing Corporation
http://www.hindawi.com
http://www.hindawi.com Volume 201-

International Journal of
International Journal of Antennas and Active and Passive Advances in
Chemical Engineering Propagation Electronic Components Shock and Vibration Acoustics and Vibration
Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014