A TWO-PHASE QUICK RESPONSE CODE TECHNIQUE FOR DOCUMENT …conferenceinfo.org/24_feb.pdf · can recover 7%, 15%, 25% and 30% of errors in the code words respectively. There have been

Conference Proceeding of

5th International Conference on Science, Technology & Management (ICSTM-2019)

Institution of Engineers, India, Sector 19A, Chandigarh, India

on 24th February 2019, ISBN: 978-93-87433-47-2

160 | Amani Jaddu, Ramakrishna S

A TWO-PHASE QUICK RESPONSE CODE TECHNIQUE FOR

DOCUMENT AND MESSAGE DISTRIBUTION

Amani Jaddu1, Ramakrishna S2

1PG Student, Department of Computer Science, Sri Venkateshwara University Tirupati 2Professor, Department of Computer Science, Sri Venkateshwara University Tirupati

Abstract

The Quick response (QR) code was intended for capacity data and fast perusing

applications. In this paper, we present another rich QR code that has two stockpiling levels and

can be utilized for report verification. This new rich QR code, named two-level QR code, has

open and private stockpiling levels. The open dimension is equivalent to the standard QR code

stockpiling level; along these lines, it is discernible by any traditional QR code application. The

private dimension is developed by supplanting the dark parts by explicit finished examples. It

comprises of data encoded utilizing q-ary code with a mistake amendment limit. This permits us

not exclusively to expand the capacity limit of the QR code, yet additionally to recognize the

first report from a duplicate. This validation is because of the affectability of the utilized

examples to the print-and-scan (P&S) process. The example acknowledgment technique that we

use to peruse the second-level data can be utilized both in a private message sharing and in a

confirmation situation. It depends on expanding the relationship esteems between P&S debased

examples and reference designs. The capacity limit can be fundamentally improved by

expanding the code letters in order q or by expanding the finished example estimate. The trial

results demonstrate an ideal reclamation of private data. It likewise features the likelihood of

utilizing this new rich QR code for report confirmation.

Keywords: Correlation, QR Code, Document Authentication, Pattern Recognition, Print-and-

scan process.

I. INTRODUCTION

Today, graphical codes, such as

EAN-13 barcode, Quick Response (QR)

code, Data Matrix, PDF417, are frequently

used in our daily lives. These codes have a

huge number of applications including

information storage (advertising, museum

art description), redirection to web sites,

track and trace (for transportation tickets or






brands), identification The popularity of

these codes is mainly due to the following

features: they are robust to the copying

process, easy to read by any device and any

user, they have a high encoding capacity

enhanced by error correction facilities, they

have a small size and are robust to

geometrical distortions.

However, those undeniable

advantages also have their counterparts:

1) Information encoded in a QR code is

always accessible to everyone, even if it

is ciphered and therefore is only legible

to authorized users (the difference

between “see” and “understand”).

2) It is impossible to distinguish an

originally printed QR code from its

copy due to their sensitivity to the Print-

and-Scan (P&S) process. In this paper,

we propose to overcome these

shortcomings by enriching the standard

QR code encoding capacity. This

enrichment is obtained by replacing its

black components by specific textured

patterns. Besides the gain of storage

capacity, these patterns can be designed

to be sensitive to distortions due to the

P&S process. These patterns that do not

introduce disruption in the standard

reading process, are always perceived as

black components by any QR code

reader. Therefore, even when the private

information is degraded or lost in the

copy, the public information is always

accessible for reading.

The proposed two level QR (2LQR)

code contains of: a first level accessible for

any standard QR code reader, therefore it

keeps the strong characteristics of the QR

code; and a second level that improves the

capacities and characteristics of the initial

QR code. The information in the second

level is encoded by using q−ary (q ≥ 2) code

with error correction capacities. This

information is invisible to the standard QR

code reader because it perceives the textured

patterns as black components. Therefore, the

second level can be used for private message

sharing. Additionally, thanks to textured

pattern sensitivity to P&S distortions, the

second level can be used to distinguish the

original 2LQR code from its copies.

This paper is organized as follows.

We start with an introduction of QR code

features and existing rich graphical codes in

Section II. In addition, the distortion added

during the P&S process will be discussed

there. The proposed two level QR (2LQR)

code as well as the proposed recognition

method are presented in Section III. In

Section IV, the experimental results show

the efficiency of the proposed recognition

methods and analyze the capacities of the






proposed 2LQR code. Finally, Section V

represents conclusions and perspectives.

The QR code was invented for the

Japanese automotive industry by Denso

Wave1 corporation in 1994. The most

important characteristics of this code are

small printout size and high-speed reading

process. The certification of QR code was

performed by International Organization of

Standardization (ISO), and its whole

specification can be found in. A QR code

encodes the information into binary form.

Each information bit is represented by a

black or a white component.

The Reed-Solomon error correction

code [15] is used for data encryption.

Therefore, one of 4 error correction levels

have to be chosen during QR code

generation. The lowest level can restore

nearly 7% of damaged information, the

highest level can restore nearly 30%. Today,

40 QR code versions are available with

different to storage capacities. The smallest

QR code version (version V1) has a 21 × 21

component size. It can store 152 bits of raw

data at the lowest correction level. The

biggest QR code version (version V40) has a

177 × 177 component size. It can store a

maximum of 7089 bits of raw data at its

lowest correction level.

As illustrated in, the QR code has a specific

structure for geometrical correction and

high-speed decoding. Three position tags are

used for QR code detection and orientation

correction. One or more alignment patterns

are used to code deformation adjustment.

The component coordinates are set by

timing patterns. Furthermore, the format

information areas contain error correction

level and mask pattern. The code version

and error correction bits are stored in the

version information areas.

The QR code generation algorithm

consists of information encoding using

Reed-Solomon error correction code,

information division on code words,

application of mask pattern, placement of

code words and function patterns into the

QR code. The QR code recognition

algorithm includes the scanning process,

image binarization, geometrical correction

and decoding algorithm.

II RELATED WORK

Image Embedding in QR Code:

The QR (Quick Response) code is a

two-dimensional barcode developed by the

Japanese company Denso-Wave in 1994,

and was approved as an ISO International

Standard and Chinese National Standard in

2000. The QR code has been widely used

due to its good features such as large data

capacity, high speed scan, and small printout

size. Increase in number of smart phones is






the reason behind popularity of QR code.

Smart phones are capable of decoding and

accessing on line resources as well as it has

high storage capacity and high speed of

decoding. QR codes are used in a various

application, such as accessing websites,

initiate phone calls, reproduce videos or

open text documents and data storing

purposes. An important problem of QR

codes is its noisy looks. To improve the

appearance of QR code and to reduce noisy

black and white random texture has

generated great interest for algorithms

capable of embedding QR codes into images

without losing decoding robustness. There

have been many efforts to improve the

appearance of such embeddings. The main

challenge of any embedding method is the

embedded result should be decodable by

standard applications. The embedding

introduces changes in the luminance of the

code, distorting the binarization thresholds

and thus increasing the probability of

detection error. The second challenge is the

problem of using the entire area of the code

in which the image or logo is to be

embedded. This cannot be done by simply

replacing information components with the

desired image. A good embedding method

should decrease the number of corrupted

components and uses the utmost area. The

proposed method is based on the selection of

a set of pixels using genetic algorithm. The

concentration of pixels and its

corresponding luminance are optimized to

minimize a visual distortion. Distortion

metric is subject to a constraint in the

probability of error.

QR code consists of black and white

square blocks called as components of a QR

code. Each component is assigned a single

bit value. Information is encoded into the

QR components. A dark component is

binary one and a light component is binary

zero. A codeword contains 8 bits of

information. There are 40 versions of QR

code. A QR code with version V has (17 +

4V) × (17 + 4V) number of components.

Therefore version 1 has 21 × 21 components

whereas version 40 corresponds to 177 ×

177 components. Fig. 1 shows the structure

of a QR code. Finder pattern contains three

identical square shape located at the three

corners of QR code. Finder pattern is the

most important pattern which enables the

detection of QR code. Alignment patterns

are also essential to locate, rotate and

aligning the QR code. Finder pattern, timing

pattern and alignment pattern are

collectively known as function pattern

region of QR code. Alignment patterns are

observed with version number 2 and

onwards however version number 1 does not

have any alignment pattern. Encoding region

within the green color consists of data and

error correction code words. Data code






words are of two types i.e. information code

words which stores the actual information

and the second is padding code words.

Encoding region stores the data, parity

components and decoding information in the

form of a code words. A codeword consist

of a block of 8 components. Quite Zone is

the guard region of QR code. QR code

utilizes RS (Reed Solomon) codes for error

correction. A QR code contains multiple RS

codes where one RS code is sufficient to

store the message. The remaining RS codes

are usually used to store non meaningful

messages [2]. There are 4 types of error

correction level i.e. L, M, Q and H which

can recover 7%, 15%, 25% and 30% of

errors in the code words respectively.

There have been a lot of efforts to improve

the appearance of QR code. The base

strategy of such work is to find the best

group of QR components to substitute by the

image or logo in the QR code. The method

presented in [3] proposed that, there are

three areas to replace the QR component by

the image or logo. These areas include data

codewords, padding codewords and the error

correcting codewords. Depending on the

error correction level of QR code, pad

characters have been changed. The size of

the embedding image in the QR code is

identified and then the image is implanted in

the identified region of QR code. The size of

the image which is to be embedded is

increased and tested the readability of QR

embedding to find largest size of which the

image could be embedded except the finder

pattern of QR code. Author concludes that if

the numbers of characters in the QR code is

decreased then the larger image can be

embedded. The second approach [12] of

embedding is based on the modification of

the pixel’s luminance. The luminance of

central pixels is modified since this is the

area usually sampled by the decoder. This

approach uses the entire area of the code for

embedding except the finder and alignment

pattern. The approach in [10] performs the

blending which combines the color image C

and the QR code Q based on the luminance

of color image and the binary value of QR

code. The blending of C and Q to produce

an output B is accomplished by replacing

pixels of Q with those of C. Author assumes

that pixels of Q are normalized so that white

pixels have a luminance of 1, and black

pixels have a luminance of 0. This algorithm

ensures that the blended output image

preserves the bright part of color image

when the pixel value of the QR code equals

to 1, and dark part of the color image when

pixel value is 0. Cox proposed a complicated

algorithm [19] to embed a binary image into

a QR code during the data encoding stage of

generating the code. He carefully

investigated the internal structure of QR

code and the logic behind data encoding,






and designed an algorithm to encode image

content as redundant numeric strings

appended to the original data. However, this

technique works only for URL type data

string and the quality of embedded image is

limited by the length of encoded URL.

Robust Message Hiding for QR Code:

Response Code (QR code) is widely

used in daily life in recent years because it

has high capacity encoding of data, damage

resistance, fast decoding and other good

characteristics. Since it is popular, people

can use it to transmit secret information

without inspection. The development of

steganography in QR code lead to many

problems arising. How to keep the original

content of QR code and embed secret

information into it are the two main

challenges. Hiding secret information based

on bit technique is so fragile to modification

attack. If an attacker changes any bit of

hidden bits, it is impossible to recover the

secret information. In this paper, we

proposed a scheme based on Reed- Solomon

codes and List Decoding to overcome this

problem. We also conduct our solution by

analyzing the complexity, security, and

experiment.

A traditional barcode is 1-dimension (1D)

barcode which only contains data by one

side. Quick Response (QR) code is a type of

2-dimension (2D) barcode developed in

1994 by Denso Wave Corporation. QR code

got this name because it was developed in

order to improve the reading speed of 2D-

barcodes. It contains data for both vertical

and horizontal dimensions. For this reason,

QR code holds a considerably greater

volume of information. It can convey

various kinds of content such as text, web

link, number, and multimedia data. The

decoding speed of the QR Code can be 20

times faster than that of other 2D symbols

[6]. In recent years, QR code is becoming

popular in business via QR readers and

mobile devices. Since QR code is so

popular, some secret information could be

transferred via it. The authors [2], [3], [4]

analyzed the properties of each QR code

before embedding it into this one. If they

want to embed a secret message into QR

code, they will encode it first. After that,

they exploit the structure of QR code which

code they want to use. It takes time, risks,

and cannot get the secret message directly

from this QR code. Lin et al. [1] observed

and proposed a novel scheme to solve this

problem. The idea to hide secret messages

into QR code is to use the error correction

capability. This idea is first proposed by Lin

et al. [1]. First of all, they encode the secret

message sm by using a shared key K and get

EK(sm). After that, they embed each bit of

EK(sm) into QR code. Their first drawback

is that if any bit of EK(sm) is damaged, it is






impossible to recover sm from QR code. The

second drawback is that if an attacker does

not change any bit of EK(sm) but adds some

extra error values into QR code, they cannot

recover their secret message. To the best of

our knowledge, all previous techniques used

bit embedding scheme to embed secret

messages into QR code. It is so vulnerable

to the modification attack, i.e. an attacker

changes any bit of secret messages. We

propose using Reed-Solomon code and List

Decoding to overcome this kind

of attack.

In Coding Theory, List Decoding, a

research field aims to correct as many errors

as possible in noisy channels, is rapidly

developed in recent years. Peter Elias [14]

and M.J. Wozencraft [15] described List

Decoding in order to correct errors over

noisy channel. Nowadays, it can be found in

many applications. It can be used to trace

who traitor is [18], [17]. Our contribution:

Our main contributions is to propose

algorithms that hide a secret message into

QR code. The secret message is invisible to

attackers and secure against modification or

damage attack. We analyze them under

complexity and security aspects, and

conduct these algorithms by experiments.

Outline of the paper: The rest of this paper is

organized as the following. Section II

presents the preliminaries. Section III

describes our proposed solution. The next

section describes the security, effectiveness,

and testing of our solution. Section V

presents experimental results. The last

section summarizes the key points and

mentions future work.

EXISTING SYSTEM

Nancy Victor proposed a technique

for data compression which enhances the

data capability of QR codes by compressing

the data previous to creation of QR codes.

B. Sklar proposed the Reed -

Solomon error correction code used for data

encryption where one of 4 error correction

levels have to be elected during QR code

generation.

R. Villán, S. Voloshynovskiy, O.

Koval, F. Deguillaume, and T. Pun proposed

the combinati on of strong text hashing and

text data hiding technologies as an effective

solution to authentication and tamper -

proofing of text documents.

T. V. Bui, N. K. Vu, T. T. P.

Nguyen, I. Echizen, and T. D. Nguyen

proposed a scheme based on reed - solomon

codes and list decoding. Using bit technique,

it hides secret information and prevents

attacker changing any bit of hidden bits.

A.E. Dirik, B. Haas discussed a copy

detection pattern tool to detect copies from

original documents and solely focus o n

counterfeit prevention.






M. Querini, A. Grillo, A. Lentini and

G.F. Italiano proposed a high capacity

colored two-dimensional code (HCC2D)

with an intention to increase barcode data

density. It supports input data of different

types and sizes and code dimension is

slickly bespoke to the real input size.

Disadvantages:

1. Storage size is high. So, it cannot save

the storage capacity.

2. It cannot restore the private information

perfectly.

3. While Compression, important

information removed.

III PROPOSED SYSTEM

We propose in this system a two-

level QR code. These two levels are public

and private level. These levels are used for

storage. The public level is the same as the

standard QR code storage level; therefore, it

is readable by any classical QR code

application. The private level is constructed

by replacing the black components by

specific textured patterns. It consists of

information encoded using q-ary code with

an error correction capacity. This allows us

not only to increase the storage capacity of

the QR code, but also to distinguish the

original document from a copy. This

authentication is due to the sensitivity of the

used patterns to the print-and-scan (P&S)

process. The pattern recognition method that

we use to read the second-level information

can be used both in a private message

sharing and in an authentication scenario. It

is based on maximizing the correlation

values between P&S degraded patterns and

reference patterns. The storage capacity can

be significantly improved by increasing the

code alphabet q or by increasing the textured

pattern size.

Advantages:

1. This proposed technique offers

significant enhancement of the data

capacity.

2. Restoration of private information is

perfectly.

3. Lossless compression with no

information lost.

IV ARCHITECTURE & SYSTEM

COMPONENTS

The system architecture is given

below and identified with the following

components.






Fig: System Architecture

Public Message Storage:

In this component, the public message is

stored in the standard QR code, using the

classical generation method. The standard

QR code generation algorithm includes the

following steps.

1. First of all, the most optimal mode

(numeric, alphanumeric, byte or Kanji)

is selected by analyzing the message

content. The message is encoded using

the shortest possible string of bits. This

string of bits is split up into 8 bit long

data code words. Then, the choice of

error correction level is performed and

the error correction codewords using the

Reed-Solomon code are generated.

2. After that, the data and error correction

codewords are arranged in the correct

order. In order to be sure that the

generated QR code can be read

correctly, the best (for encoded data)

mask pattern is applied.

3. After this manipulation, the codewords

are placed in a matrix in a zigzag

pattern, starting from the bottom-right

corner. The final step is to add the

function patterns (position tags,

alignment, timing, format and version

patterns) into the QR code.

Private Message Encoding:

1. In this component, the private row-bit

string is encoded using error correction

code (ECC) to ensure the message error

correction after the P&S operation. We

use the block codes, and more precisely

cyclic codes (or polynomial-generated

codes) such as Golay code or Reed-

Solomon code, for message encoding.

2. The public level is identical to the

standard QR code storage level, read by

any classical QR code application

whereas the private level is made by

replacing the black components by

specific textured patterns.






Black Component Replacement:

1. In 2LQR code black and white

components are represented using zeros

and ones. Cell is divided into 24x24

pixel size.

2. Check for zeros and whole of the zeros

will be replaced with code. The textured

pattern which replaces the black

components is based on the number of

zeros available.

For example, if there are 5 zeros then 5

squares corresponding to that will be drawn

while encoding. During decoding the same 5

squares will be decoded as 5.

V CONCLUSION

In this paper a new rich code called

two level QR (2LQR) code is proposed. This

2LQR code has two levels: a public level

and a private level. The public level can be

read by any QR code reading application,

while the private level needs a specific

application with specific input information.

This 2LQR code can be used for private

message sharing or for authentication

scenarios. The private level is created by

replacing black modules with specific

textured patterns. These textured patterns are

considered as black modules by standard QR

code reader. Thus, the private level is

invisible to standard QR code readers. In

addition, the private level does not affect in

anyway the reading process of the public

level. The proposed 2LQR code increases

the storage capacity of the classical QR code

due to its supplementary reading level.

Experiment results show that the storage

capacity is improved by up to 28%

(transition from message size equal to 272

bits to a message length of 380 bits). The

storage capacity of the2LQR code can be

improved by increasing the number of

textured patterns used or by decreasing the

textured pattern size. All experiments show

that even with a pattern size of 6×6 pixels

and with an alphabet dimension q = 8, it is

possible to obtain good pattern recognition

results, and therefore a successful private

message extraction. However, we are facing

a trade-off between the pattern size, the

alphabet dimensions and the quantity of

stored information during the 2LQR code

generation. One important feature of the

textured patterns used is their sensitivity to

the P&S process. To take advantage of this

sensitivity, we use a pattern recognition

method based on maximization of

correlation values among the P&S degraded

versions and characterization patterns. We

have tried three different types of

characterization patterns: mean patterns,

median patterns (for the private message

sharing scenario) and original patterns (for

the document authentication scenario). The

mean and median characterization patterns

give almost the same results of pattern






detection. Therefore, either of them can be

used in the private message sharing

scenario. The best pattern recognition results

were obtained, when the original patterns

are used as characterization patterns. The

original patterns can be also used for the

private message sharing scenario, but in this

case the blind method for pattern detection

cannot be performed. The suggested

textured patterns can be distinguished only

after one P&S process. Therefore, we can

use the detection method with original

patterns in order to ensure good document

authentication results.

FUTURE WORK:

In our future work, we will address

five different paths. The first path will

concern the improvements of the pattern

recognition method. The second will cover

the textured pattern analysis to automate its

combination process. The third will deal

with message recovering and authentication

attacks, such as cropping and code

reconstruction. The forth path will concern

the study of the second level recovery

problems in the 2LQR code images captured

by a camera. In the last path, the storage

capacity of 2LQR code will be increased by

replacing also the white modules with

textured patterns, which have small density

than black pixels.

REFERENCES

[1] Information Technology Automatic

Identification and Data Capture

Techniques EAN/UPC Bar Code

Symbology Specification, ISO/IEC

Standard 15420:2009, 2009.



Techniques—Data Matrix Bar Code

Symbology Specification, ISO/IEC

Standard 16022:2006, 2006.



Techniques—Bar Code Symbology—

QR Code, ISO/IEC Standard

18004:2000, 2000.

[4] Z. Baharav and R. Kakarala, “Visually

significant QR codes: Image blending

and statistical analysis,” in Proc. IEEE

Int. Conf. Multimedia Expo (ICME),

Jul. 2013, pp. 1–6.

[5] C. Baras and F. Cayre, “2D bar-codes

for authentication: A security

approach,” in Proc. 20th Eur. Signal

Process. Conf. (EUSIPCO), Aug.

2012, pp. 1760–1766.

[6] T. V. Bui, N. K. Vu, T. T. P. Nguyen,

I. Echizen, and T. D. Nguyen, “Robust

message hiding for QR code,” in Proc.

IEEE 10th Int. Conf. Intell. Inf. Hiding

Multimedia Signal Process. (IIH-

MSP), Aug. 2014, pp. 520–523.






[7] A. T. P. Ho, B. A. M. Hoang, W.

Sawaya, and P. Bas, “Document

authentication using graphical codes:

Reliable performance analysis and

channel optimization,” EURASIP J.

Inf. Secur., vol. 2014, no. 1, p. 9,2014

[8] T. Langlotz and O. Bimber,

“Unsynchronized 4D barcodes,” in

Proc. 3rd Int. Symp., ISVC 2007,

Lake Tahoe, NV, USA, Nov. 26–28,

2007, pp. 363–374.

[9] C.-Y. Lin and S.-F. Chang, “Distortion

modeling and invariant extraction for

digital image print-and-scan process,”

in Proc. Int. Symp. Multimedia Inf.

Process., 1999, pp. 1–

[10] P.-Y. Lin, Y.-H. Chen, E. J.-L. Lu, and

P.-J. Chen, “Secret hiding mechanism

using QR barcode,” in Proc. IEEE Int.

Conf. Signal-Image Technol. Internet-

Based Syst. (SITIS), Dec. 2013, pp.

22–25.

[11] J. Picard, “Digital authentication with

copy-detection patterns,” Proc. SPIE,

vol. 5310, pp. 176–183, Jun. 2004.

[12] M. Querini, A. Grillo, A. Lentini, and

G. F. Italiano, “2D color barcodes for

mobile phones,” Int. J. Comput. Sci.

Appl., vol. 8, no. 1, pp. 136–155,

2011.

[13] M. Querini and G. F. Italiano, “Facial

biometrics for 2D barcodes,” in Proc.

IEEE Fed. Conf. Comput. Sci. Inf.

Syst. (FedCSIS), Sep. 2012, pp. 755–

762.

[14] J. Rouillard, “Contextual QR codes,”

in Proc. IEEE 3rd Int. Multi-Conf.

Comput. Global Inf. Technol.

(ICCGI), Jul./Aug. 2008, pp. 50–55.

[15] B. Sklar, Digital Communications,

vol. 2. Englewood Cliffs, NJ, USA:

Prentice-Hall, 2001.

[16] K. Solanki, U. Madhow, B. S.

Manjunath, S. Chandrasekaran, and I.

El-Khalil, “‘Print and scan’ resilient

data hiding in images,” IEEE Trans.

Inf. Forensics Security, vol. 1, no. 4,

pp. 464–478, Dec. 2006.

[17] M. Sun, J. Si, and S. Zhang, “Research

on embedding and extracting methods

for digital watermarks applied to QR

code images,” New Zealand J.

Agricult. Res., vol. 50, no. 5, pp. 861–

867, 2007.

[18] I. Tkachenko, W. Puech, O. Strauss, J.-

M. Gaudin, C. Destruel, and C.

Guichard, “Fighting against forged

documents by using textured image,”

in Proc. 22th Eur. Signal Process.

Conf. (EUSIPCO), Sep. 2014, pp.

790–794.

[19] R. Ulichney, Digital Halftoning.

Cambridge, MA, USA: MIT Press,

1987.

[20] R. Villán, S. Voloshynovskiy, O.

Koval, F. Deguillaume, and T. Pun,






“Tamper-proofing of electronic and

printed text documents via robust

hashing and data-hiding,” in Proc.

SPIE, vol. 6505, p. 65051T, Feb.

2007.

Amani Jaddu she is a master

of Computer Science (M.Sc)

pursuing in Sri Venkateswara

University, Tirupati, A.P. She

received Degree of Bachelor of

Science in 2017 from Rayalaseema

University, Kurnool. Her research interests

are Cloud Computing, Data Warehousing,

and Big Data.





173 | Aruna Kumari Yanduri, Ramakrishna S

An Artificial Intelligence Technique for Explorated Searchable

Query Processing

Aruna Kumari Yanduri1, Ramakrishna S2 1PG Student, Department of Computer Science, Sri Venkateshwara University Tirupati

2Professor, Department of Computer Science, Sri Venkateshwara University Tirupati

Abstract

Exploratory search is an increasingly imperative movement for Web searchers. Be that as

it may, the ebb and flow seek framework cannot give adequate help to exploratory inquiry. In

this manner, we made inside and out investigation for exploratory pursuit procedures, and found

that there are a ton of hunt objective move marvels in exploratory inquiry. In light of this reality,

we have planned another inquiry proposal technique to help exploratory pursuit. Right off the

bat, as per the social qualities of searchers in the inquiry objective move forms, every one of the

questions submitted in the pursuit objective move forms are extricated from web crawler logs

utilizing AI. And after that we have utilized the inquiries to assemble a pursuit objective move

diagram; at long last, the arbitrary walk calculation is utilized to get the inquiry suggestions in

the hunt objective move chart. Likewise, we showed the viability of the strategy for exploratory

pursuit by contrasting examinations and alternate strategies.

Keywords: Exploration Search, Query Recommendation, Artificial Intelligence.

I. INTRODUCTION

Exploratory search is an increasingly

important activity yet challenging for Web

searchers. In exploratory search, the

searcher is unfamiliar with their problem

domain, ensures about the ways to achieve

their goal, or lacks a well-defined goal. To

support exploratory search, the search

system is required not only to provide

accurate search results, but also to help

searchers explore related and novel aspects.

Therefore, exploratory search system needs

an effective query recommendation method

to re-solve this problem.

However, the current query

recommendation methods mainly focus on

optimizing users’ current query which is far

away from satisfying users’ information

needs of the whole search session. To

support exploratory search, we observed and

analyzed the search logs of exploratory

search process performed by different users,

and we found that there are a lot of search

goal shift phenomena in exploratory search.

As the following example: A Chinese

university student attends a birthday party

organized by a French student, and he wants

to choose a suit-able birthday gift, which is a

typical exploratory search task. Because the






Chinese student only got some very vague

goals, such as

Object: a gift not a normal thing

Applicable occasions: birthday party

Basic features: French favorite items

Budget: 200 RMB or so

Based on these conditions, the

student used the key words “French people

like flowers” for the query; explored

“flowers” which is the most popular gift. He

felt using flowers as a birthday gift doesn’t

feature after clicking many links of search

results. And the search results mentioned

that French people are very fond of drinking

wine. So, he changed his idea and felt that

“wine” may be more appropriate as a gift for

the birthday party. So, the user used “French

wine” as a key word and query “red wine”

as new search goal to explore. Using the

search results about the “French wine brand”

and “French wine prices”, the student

figured out French wine prices are expensive

far beyond his budget. Obviously “red wine”

is not a suitable search goal either. At the

same time, he thought, “arts and crafts” may

be more appropriate. Then he used

“handicrafts”, “Chinese arts and crafts” as

key words to query on the "arts and crafts"

which is a new search goal, and eventually

found hopeful gift to the end of the search

task.

From the example, it’s clear that the

user's search goal shifts from the “flowers”

to “red wine” and then from the “wine” to

“arts and crafts”. And the search goal shifts

precisely reflect the user’s exploratory

behaviors and needs. Therefore, we based on

the "search goal shift" de-signed a new

recommendation method to support

exploratory search. Firstly, according to the

user’s behavioral characteristics in the

search goal shift process, we extracted all

queries during search goal shift processes

from search logs; then we used the queries

to construct a search goal shift graph;

finally, we recommended other goals related

to the current goals using the search goal

shift graph.

In addition, we have designed a

query recommendation test method, by

which we can compare our recommendation

method with the other methods. And the

experimental results showed that the

recommendation method we designed can

significantly shorten the search.

II RELATED WORK

Query Recommendation

Most of the query recommendation

techniques are using similarity measures

between queries by query terms, clicked

documents, or sequences of queries in

sessions. Baeza-Yates et al. [2] extracted

query-clicked URL/doc bipartite graphs

using search logs to find query

recommendations. Craswell and Szummer

[3] also used the query-click graph to find

related documents and queries. Mei et al. [4]

presented a “Hitting Time” algorithm to find

related queries using the query-click graph.

Cao et al. [5] tried to understand user's

context which in-cludes multiple






information including age, gender,

username, IP, tools etc. and also previous

queries in a query session in order to suggest

new queries. Boldi et al. [6] proposed a

query-flow graph which represents the latent

querying behavior contained in a query log.

Exploratory Search

In the past 30 years, many scholars

have made in-depth study of the search

process of exploratory search behavior. In

1989, Dr. Bates M J proposed Berry picking

model [7] that the user's search direction and

the desired result will constantly change

with the search process changing. In 1991,

Kuhlthau C C proposed that information

retrieval process includes starting, selection,

exploration, collection and ending six stages

[8]. In 1995, Byström K and Järvelin K used

the methods of logs and questionnaires to

analyze the relationship with search

complexity of the task, type of information,

information channels and resources [9]. In

2006, Marchionini G proposed exploratory

search [10].

Recently, exploratory search

research focuses on the characteristics of the

exploratory search process and the different

types of support needed to help people make

exploratory searches [1]. Someone tries to

provide a query preview control by allowing

users to take nodes and rec-ord the results

[11] so that they can view the distribution of

newly-retrieved and retrieved documents

before running the query [12]. Some

research efforts focus on traditional search

techniques such as query suggestions,

aspects and information classification. For

example, Hassan Awadallah et al. [13]

constructed a method of automatically

identifying and recommending tasks that

allow searchers to explore and complete

complex search tasks, Sun et al. [14]

proposed a topic-oriented query for explor-

atory search method, Ksikes et al. [15]

designed an ex-ploratory faceted search

system, Zhang et al. [16] grouped the

relationships between entities into a virtual-

generated hierarchical clustering to an

effective leader to explore and discover.

Other attempts have been made to design

and research visual search interfaces and

interactive user modeling to support

exploratory search tasks. For exam-ple,

Bron et al. [17] proposed an auxiliary

exploratory search interface to support

media research; Bespinyowong et al [18]

designed exploratory data ex-ploratory

ranking interface; Peltonen et al [19] used a

negative feedback search intent radar

interface to help users conduct exploratory

search.

All these previous methods focus on

refining user requirements, showing several

facets helping users refine their requirement

and find their desired information. But they

cannot satisfy such user needs as finding

some novel search goals when users are

losing the interests of current search goal.

EXISTING SYSTEM

The current query recommendation

methods mainly focus on optimizing users’

current query which is far away from






satisfying users’ information needs of the

whole search session. To support

exploratory search, we observed and

analyzed the search logs of exploratory

search process performed by different users,

and we found that there are a lot of search

goal shift phenomena in exploratory search.

Most of the query recommendation

techniques are using similarity measures

between queries by query terms, clicked

documents, or sequences of queries in

sessions. In existing system, they used

extracted query-clicked URL/doc bipartite

graphs using search logs to find query

recommendations.

Disadvantages:

All these previous methods focus on

refining user requirements, showing several

facets helping users refine their requirement

and find their desired information. But they

cannot satisfy such user needs as finding

some novel search goals when users are

losing the interests of current search goal.

III PROPOSED SYSTEM

This paper presents the search goal

shifts precisely reflect the user’s exploratory

behaviors and needs. Therefore, we based on

the "search goal shift" de-signed a new

recommendation method to support ex-

ploratory search. Firstly, according to the

user’s behavior-al characteristics in the

search goal shift process, we extracted all

queries during search goal shift processes

from search logs; then we used the queries

to construct a search goal shift graph;

finally, we recommended other goals related

to the current goals using the search goal

shift graph. Based on the basic framework of

the search goal shift graph, exploratory

query recommendation method mainly

consists of two parts, offline and online.

Our final goal is to provide query

recommendations for users, the process of

“identifying search goal shift” is to identify

all search goal shift query pairs from the

search en-gine logs and use them to

compose of the candidate set.

IV SYSTEM ARCHITECTURE

Based on the basic framework of the

search goal shift graph, exploratory query

recommendation method mainly consists of

two parts, offline and online.

Fig: System Architecture






Offline Part

Offline part mainly includes two

major steps, the search goal shift

identification and the search goal shift graph

building. In the offline part, we manually

annotate the search goal shift in some users’

search session, then use machine learn-ing

to convert inefficient manual identification

process into efficient AI calculation. Finally,

we use all queries submitted during search

goal shifts to construct a search goal shift

graph.

Online Part

Online part also contains two steps,

user's search behavior judgment and top-k

recommend. In the online part, we use the

identification model which is trained from

the offline part to judge whether users’

search behaviors belong to “search goal

shift”, then we use a random walk algorithm

to find the top-k most relevant search

queries from the search goal shift graph as a

result of recommendation.

V CONCLUSION

In this paper, we contemplated the

pursuit objective move which is one of the

imperative conduct attributes of exploratory

inquiry, and planned another question

suggestion technique dependent on the hunt

objective move to help exploratory hunt.

The strategy utilizes AI to uncover all

inquiries amid pursuit objective move forms

from internet searcher logs to assemble the

inquiry objective move chart, and uses

irregular walk calculation to acquire

question suggestions in the hunt objective

move diagram. In the meantime, we

demonstrated the adequacy of the suggestion

strategy by the relative investigations with

different techniques.

Future Enhancement:

It is not possible to develop a system

that makes all the requirements of the user.

User requirements keep changing as the

system is being used. Some of the future

enhancements that can be done to this

system are:

As the technology emerges, it is

possible to upgrade the system and can be

adaptable to desired environment.

Based on the future security issues, security

can be improved using emerging

technologies like single sign-on.

VI REFERENCES

[1] White R W, Roth R A. “Exploratory search: Beyond the query response

paradigm,” Synth Lect Inf Concept Retr

Serv, vol. 1, no. 1, pp. 1-98, 2009.

[2] Baeza-Yates R, Hurtado C, Mendoza M. “Query recommendation using query

logs in search engines,” in Proc.

International Conference on Extending

Database Technology, pp. 588-596,

2004.

[3] Craswell N, Szummer M. “Random walks on the click graph,” in Proc. The






30th annual international ACM SIGIR

conference on Research and

development in information retrieval, pp.

239-246, 2007.

[4] Cao H, Jiang D, Pei J, et al. “Context-aware query suggestion by mining click-

through and session data,” in Proc. The

14th ACM SIGKDD international

conference on Knowledge discovery and

data mining, pp.875-883, 2008.

[5] Mei Q, Zhou D, Church K. “Query suggestion using hitting time,” in Proc.

The 17th ACM conference on

Information and knowledge

management, pp. 469-478, 2008.

[6] Boldi P, Bonchi F, Castillo C, et al. “Query suggestions using query-flow

graphs,” in Proc. The 2009 workshop on

Web Search Click Data, pp. 56-63, 2009.

[7] Bates M J. “The design of browsing and berry picking techniques for the online

search interface,” Online review, vol. 13,

no. 5, pp. 407-424, 1989.

[8] Kuhlthau C C. “Inside the search process: Information seeking from the

user's perspective.” Journal of the

American society for information

science, vol. 42, no. 5, pp. 361-424,

1991.

[9] Byström K, Järvelin K. “Task complexity affects information seeking

and use.” Information processing &

management, vol. 31, no. 2, pp. 191-213,

1995.

[10] Marchionini G. “Exploratory search: From finding to under-standing.”

Communications of the ACM, vol. 49,

no. 4, pp. 41-46, 2006.

[11] Donato D, Bonchi F, Chi T, et al. “Do you want to take notes? identifying

research missions in Yahoo! search

pad,” in Proc. The 19th ACM

international conference on World Wide

Web, pp. 321-330, 2010.

[12] Qvarfordt P, Golovchinsky G, Dunnigan T, et al. “Looking ahead:

query preview in exploratory search,” in

Proc. The 36th international ACM

SIGIR conference on Research and

devel-opment in information retrieval,

PP. 243-252, 2013.

[13] Hassan Awadallah A, White R W, Pantel P, et al. “Supporting complex

search tasks,” in Proc. The 23rd ACM

International Conference on Conference

on Information and Knowledge

Management, pp. 829-838, 2014.

[14] Sun H C, Jiang C J, Ding Z J, et al. “Topic-Oriented Exploratory Search

Based on an Indexing Network.” IEEE

Transactions on Systems, Man, and

Cybernetics: Systems, vol. 46, no.2, pp.

234-247, 2016.

[15] Ksikes, A. “Towards exploratory faceted search systems.” Doctoral

dissertation, University of Cambridge,

2014

[16] Zhang Y, Cheng G, Qu Y. “Towards exploratory relationship search: A

clustering-based approach,” in Proc.

Joint International-al Semantic

Technology Conference. Springer

International Publishing, pp. 277-293,

2013.

[17] Bron M, Van Gorp J, Nack F, et al. “A subjunctive exploratory search interface






to support media studies researchers,” in

Proc. ACM SIGIR conference on

Research and development in

information retrieval, pp. 425-434, 2012.

[18] Bespinyowong R, Chen W, Jagadish H V, et al. “ExRank: an exploratory

ranking interface.” Proceedings of the

VLDB Endowment, vol. 9, no. 13,

pp.1529-1532, 2016.

[19] Peltonen J, Strahl J, Floréen P. “Negative Relevance Feedback for

Exploratory Search with Visual

Interactive Intent Modeling,” in Proc.

The 22nd ACM International

Conference on Intelligent User

Interfaces, pp.149-159, 2017.

ARUNA KUMARI

YANDURI she is a master

of Computer Science (M.Sc)

pursuing in Sri

Venkateswara University,

Tirupati, A.P. She received

Degree of Bachelor of Science in 2017 from

Rayalaseema University, Kurnool. Her

research interests are Artificial Intelligence,

Machine Learning, and Quantum

Computing.





180 | Hari Varma B, Anjan Babu G

A Deep Learning Enhanced Technique for Classification of Blood

Cell Images

Hari Varma B1, Anjan Babu G2

1PG Student, Department of Computer Science, Sri Venkateshwara University Tirupati 2Professor, Department of Computer Science, Sri Venkateshwara University Tirupati

Abstract

The problem of identifying and counting blood cells within the blood smear is of both theoretical

and practical interest. The differential counting of blood cells provides invaluable information to

pathologist for diagnosis and treatment of many diseases. In this paper we propose an efficient

hierarchical blood cell image identification and classification method based on multi-class

support vector machine. In this automated process, segmentation and classification of blood cells

are the most important stages. We segment the stained blood cells in digital microscopic images

and extract the geometric features for each segment to identify and classify the different types of

blood cells. The experimental results are compared with the manual results obtained by the

pathologist, and demonstrate the effectiveness of the proposed method.

Keywords: Artificial Intelligence, Convolutional Neural Network, Machine Learning.

I. INTRODUCTION

It is notable that platelets principally

incorporate red blood cells, white platelets

and platelets. In blood, leucocyte assumes an

essential job in the human insusceptible

capacity, so it is likewise called the safe cell.

More often than not, hematologists use

granulated data and shape data in leukocytes

to isolate white platelets into granular cells:

neutrophil, eosinophil, basophil and non-

granular cells: monocyte and lymphocyte.

The extent in the blood of these five types

of cells is diverse for the sick and non-ailing

bloods. Specialists regularly utilize this

fundamental information as criteria for

deciding the sort and seriousness of this

ailment. In this manner, the investigation of

white platelet classification has critical

significance and esteem for restorative

conclusion. In light of the significance of

platelet classification in the conclusion,

scientists have proposed numerous

calculations to order platelets. In 2003, Sinha

and Ramakrishnan [1] classified cells

utilizing SVM with an acknowledgment rate

of 94.1%. In 2006, Yampri et al. [2] utilized






100 pictures to play out the same trials. They

actualized the programmed edge also,

versatile shape to fragment cells, and utilized

the littlest blunder strategy to group them,

and the acknowledgment rate was 96% [2].

Yampri et al. [2] used the KNN calculation.

Be that as it may, the KNN calculation does

not deal with uneven tests well. In the event

that the example limit of a class is vast, while

the example limit of different classes is little,

a few issues emerge. For instance, when

another example is contribution to the

analytic framework, it might result in a class

with a vast limit of being overwhelming in

the K closest neighbors of this example.

What's more, the calculation is

computationally costly on the grounds that

each example should be sorted in request to

compute its separation from every single

referred to test so as to get its K closest

neighbors.

II RELATED WORK

Previously related blood cell

classification algorithms mainly include the

KNN algorithm, Bayesian classifier, SVM

classifier, etc. We briefly review and discuss

in this section. The core idea of the KNN

algorithm is that if most of the k most

adjacent samples in a feature space belong to

a certain category. Note that the sample also

has the characteristics of all the other samples

in this category. This method determines

the class in which the sample is to be

classified based on the category of the nearest

samples in determining the classification

decision. The KNN method is only relevant

to a very small number of neighboring

samples in the category decision. Based on

this theory, Young (1972) experimented with

199 cell images. He first used histogram

thresholds to segment white blood cells and

classified them using a distance classifier.

The recognition rate was 92.46% [24]. Bikhet

et al. [25] used entropy based and iterative

thresholding methods to divide cells and

classify them with a distance classifier, with

a recognition rate of 90.14%. Bayesian

classification is based on statistical

classification and uses its knowledge of

probability statistics to classify data. In many

classifications, naive Bayes algorithm can

be compared with decision tree and neural

network algorithm. Sinha and Ramakrishnan

[1] used Bayesian classifiers to classify cells

and the recognition rate was 82.3%. The era-

Umpon and Dhompongsa (2007) used a

Bayesian classifier to classify the bone

marrow images of the Ellis Fisher Cancer

Center at the center of Missouri (only one

cell per picture), and the recognition rate was

77% [26], [27]. Ghosh et al. [28] used a

watershed algorithm to segment 150 cell

images and classify them using a Bayesian

classifier, and the recognition rate was

83.2%. The classification idea of SVM is

essentially similar to the linear regression LR






classification method. It is to obtain a set of

weight coefficients that can be classified after

linear representation. SVM _rst trains a

separation hyper-plane, and then the plane is

the decision boundary of the classification.

Classical SVM algorithm is only suitable for

two types of classification problems. After

improvement, SVM can also be applied to

multiple classification problems. In the

actual application of white blood cell

classification, it is generally necessary to

solve the problem of multiple classifications.

For example, the five-classification problem

of leukocytes we studied can be solved by

combining multiple binary SVM. Rezato_ghi

and Soltanian-Zadeh [29] used the Gram-

Schmidt Orthogonal and Snake algorithm to

segment 400 blood smears and classified

them using SVM. Their recognition rate was

90% [29]. Recently, convolutional neural

networks have been widely implemented in

various image classification fields. In

particular, convolutional neural networks

(ConvNets) [11] achieved unprecedented

results in the 2012 ImageNet large-scale

visual recognition challenge, which included

classifying natural images in the ImageNet

dataset into 1000 _ne-grained categories [3].

They also significantly improve the

performance of various medical imaging

applications [30], [31], such as classification

of lung diseases and lymph nodes in CT

images [32], [33], segmentation (pixel

classification) of brain tissues in MRI [34],

vessel segmentation based on fundus images

[37], and detecting cervical intraepithelial

neoplasia (CIN, particularly CIN2C) at

patient level based on Cervi gram images or

Multimodal data [36]. In addition, ConvNets

showed superior performance in cell image

classification such as pleural cancer [38] and

human epithelial cell images [39]. Although

these methods can be used to generate good

classification engines, they still have some

drawbacks. Traditional machine learning

methods (such as SVM) need to extract

features manually. The acquisition of features

mainly depends on the designer's prior

knowledge. This feature extraction method is

difficult to make full use of the information

contained in the image, and will increase the

designer's workload. The deep learning

algorithm effectively solves this problem. It

can automatically learn the effective features

of the image. Deep learning algorithms such

as deep residual network also have good

performance in image classification tasks.

However, these neural network classification

algorithms cannot fully utilize some features

of the image that have a long-term

dependency relationship with image labels,

and thus these classification methods cannot

classify cell images like people with memory.

For this purpose, we introduce a recurrent

neural network and fuse it with a

convolutional neural network to perform the

task of blood cell image classification.






III PROPOSED SYSTEM

Convolutional Neural Networks

Both the 2-dimensional and 3-dimensional

structures of an organ being studied are

crucial in order to identify what is normal

versus abnormal. By maintaining these local

spatial relationships, CNNs are well-suited to

perform image recognition tasks. CNNs have

been put to work in many ways, including

image classification, localization, detection,

segmentation and registration. CNNs are the

most popular machine learning algorithm in

image recognition and visual learning tasks,

due to its unique characteristic of preserving

local image relations, while performing

dimensionality reduction. This captures

important feature relationships in an image

(such as how pixels on an edge join to form a

line), and reduces the number of parameters

the algorithm has to compute, increasing

computational efficiency. CNNs are able to

take as inputs and process both 2-

dimensional images, as well as 3-dimensional

images with minor modifications. This is a

useful advantage in designing a system for

hospital use, as some modalities like X-rays

are 2-dimensional while others like CT or

MRI scans are 3-dimensional volumes.

CNNs and Recurrent Neural Networks

(RNNs) are examples of supervised machine

learning algorithms, which require significant

amounts of training data. Unsupervised

learning algorithms have also been studied

for use in medical image analysis. These

include Autoencoders, Restricted Boltzmann

Machines (RBMs), Deep Belief Networks

(DBNs), and Generative Adversarial

Networks (GANs)

Fig 3 : The Flow Chart of automatic

recognition of blood cells

IV METHODOLOGY

Colour Split Channel

The blood smear may be stained by

different colour dyes. To avoid being

influenced by dye colour, all blood smear

images were first transformed into gray level.

A typical peripheral blood smear image

consists of four components, which are the

background, erythrocytes, leukocytes, and

thrombocytes. Leukocytes appear rather






darker than the background, and erythrocytes

appear in an intermediate intensity level. To

segment the desired object from the

background, it is found that the green

component of the RGB input image gives the

best contrast between the background and the

blood cells components, as shown in Fig. 4.

As a result, the green channel is used to

segment the blood cells in our proposed

method.

Image Segmentation

Image segmentation consists basically on

partitioning an image into a set of disjoint

and homogeneous regions which are

supposed to correspond to image objects that

are meaningful to a certain application. Thus,

the segmentation process is based on using

thresholding, morphology, and watershed to

enclose every element in the blood slide in a

distinct area.

Binary

In order to segment the desired object from

the background, we need to generate a binary

image that separates foreground and

background image pixels. To produce a

representative binary image, Otsu’s adaptive

threshold algorithm [7] is applied on the

green channel to classify all the pixels into

two classes. Otsu’s method exhaustively

searches for the threshold Tc that minimizes

the within-class variance, defined as a

weighted sum of variances of two classes:

Where the weight pi is the probability of a

pixel in the i-th class separated by a threshold

t and the variance of pixels’ gray level

intensities in the i-th classes. Fig. 5 shows the

output binary image produced corresponding

to that shown in Fig. 4(c).

Figure 5 The binary blood cell image

generated by Otsu’s threshold algorithm.

Mathematical Morphology






Mathematical morphology operations

[8] are nonlinear, translation invariant

transformations. The basic morphological

operations involving an image S and a

structuring element E are

where denote the set intersect respectively. E

+ s denotes the translation of a s. The

opening and closing derived from dilation are

defined by Mathematical morphology

operations are holes in blood cells and to

remove the unwanted blood cells and

background. Watershed

The objective of watershed

segmentation of the highest gray levels,

which are called the simplest way to explain

watershed segmentation approach.” Imagine

that a hole minimum of the surface, and we

flood was catchment basins from the holes. If

the w catchment basins are likely to merge

due to fu a dam is built to prevent the

merging. This will eventually reach a stage

when only the watershed lines) is visible

above the water order to separation of

overlapping cells, water is applied on

distance transform of binary having larger

area. Fig. 6 shows the watershed result for

the blood cell image.

Feature Extraction

MATERIALS AND METHODS

stages Image Data Collection: The blood

specimens were obtained from different

patients with sickle cell anaemia, sickle cell

disease and normal volunteers.

Each blood cell image contains

number of normal and abnormal cells. Blood

Cell Segmentation: Image segmentation is

used to detect the entire blood cells

(Dougherty, 1994; Wroblewska et al., 2003).

in a segmented image, the picture elements

are no longer the pixels, but connected set of

pixels, all belonging to the same region. An

object can be easily detected in an image if

the object has sufficient contrast from the

background. We use edge detection and basic

morphology tools to detect a cell.






The individual cells are close to each

other and the borders among them are not

well defined. The morphological operations

aim at extracting relevant structures of the

image by probing the image with another set

of a known shape called structuring element,

chosen as the result of prior knowledge

concerning the geometry of the relevant and

irrelevant image structures.

The most known morphological operations

include erosion, dilation, opening and

closing. The morphological approach to

image segmentation combines regions

growing and edge detection techniques

(Serra, 1984; Ponsen et al., 2009). The

applied procedure of the image segmentation

and cell separation consists of the following:

Transformation of the original image into

gray scale. Detect the entire cell using edge

detection technique Application of dilation

and erosion operations to smooth the object

and to eliminate the distortions Feature

Extraction: Twenty-seven features were

extracted from each cell image (Table 1).

This included 4 geometrical features, 16

statistical features and 7 moment invariant

features (Osowski et al., 2004; Santinelli et

al., 2002). Geometrical Features Description:

We use the following geometrical features to

study characteristics of the cells:

Area A-the number of pixels on the interior

of the cell

Perimeter P-the total distance between

consecutive points of the border

Compactness C-given by the formula:

perimeter2 /area • Form factor F-

4*3.14*Area/Perimeter2

SVM Classification

Support vector machine (SVM) [10] is a

concept for a set of related supervised

learning methods that analyze data and

recognize patterns, used for classification and

regression analysis. The main advantage of

the SVM network used as a classifier is its

very good generalization ability and

extremely powerful learning procedure,

leading to the global minimum of the defined

error function. Given instances xi, i=1, …, l

with labels 8$! K2K3, the main task in

training SVMs is to solve the following

quadratic optimization problem [11]:

where e is the vector of all ones, C is the

upper bound of all variables, Q is an l by l

symmetric matrix with Qij = yiyjK(xi, xj),

and K(xi, xj) is the kernel function. The most

known kernel functions are the radial

Gaussian basis, polynomial, spline, or

sigmoidal functions. The final learning

problem of the SVM is transformed to the

solution of the so-called dual problem

defined with respect to the Lagrange

multipliers [12]:






where b is the bias and the vector x represent

the class when s(x) is positive and the

alternative class when s(x) is negative. The

hyperparameter of the kernel function and the

regularization constant C have been adjusted

by repeating the learning experiments for the

set of their predefined values and choosing

the best value on the validation data sets.

Their optimal values are those for which the

classification error on the validation data set

was the smallest.

The one-against-one method [13] is

applied to deal with the problem of multiple

classes. The maximum voting of the multiple

classes is used to find the final classification

results. During the training phase, the models

of the multiple classes SVMs are learned

from training data. In the testing phase, the

learned models are employed to generate

multiple sets of predictions for each test

sample. The one having the largest prediction

is the final decision.

From earlier literatures, we found that

it is hard to accurately distinguish blood cells

into seven classes by using the single-stage

SVM classification. Thus, we propose the

hierarchical SVM classification to improve

the recognition ratio. Fig. 8 illustrates our

proposed hierarchical strategy. For fast and

efficient classification, five features, area,

histogram, circularity, cytoplasm ratio, and

color of cytoplasm, are extracted for the

following SVM training. For the first level,

blood cells can be distinguished into two

types, thrombocytes and erythrocytes,

leukocytes by the feature “area.” Next, we

can use the feature “histogram” to identify

erythrocytes and leukocytes. For leukocytes,

we can use the feature “circularity” to

identify granulocytes and agranulocytes due

to agranulocytes belong to the mononuclear

cell group. In the following, we use the

feature “color of cytoplasm” to distinguish

granulocytes into neutrophils, eosinophils,

and basophils. Finally, monocytes and

lymphocytes can be recognized by the feature

“cytoplasm ratio.”






V CONCLUSION

This study demonstrated an efficient

hierarchical blood cells classification method

using the geometric features from the nucleus

and the cytoplasm and a multi-class SVM

classification scheme. Classification using

the proposed hierarchical strategy

outperformed classification using only the

single-stage SVM because the cytoplasm of

some leukocytes presents a very weak

difference against the background and

touches neighboring cells. In addition,

experimental results showed that using the

hierarchical multi-class SVM classification

with hierarchical features could indeed

improve the classification performance

compared to the sing

Documents

A TWO-PHASE QUICK RESPONSE CODE TECHNIQUE FOR DOCUMENT …conferenceinfo.org/24_feb.pdf · can recover 7%, 15%, 25% and 30% of errors in the code words respectively. There have been