artificial neural network based optical.pptx

17
8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 1/17 ARTIFICIAL NEURAL NETWORK BASED OPTICAL CHARACTER RECOGNITION

Upload: preyassharma

Post on 06-Jul-2018

216 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 1/17

ARTIFICIAL NEURAL NETWORK BASED OPTICAL

CHARACTER RECOGNITION

Page 2: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 2/17

  BSTR CT

1. Optical Character Recognition deals in recognition and classification of character 

s from an image. For the

2. recognition to be accurate, certain topological and geometrical properties are c

alculated, based on which

3. a character is classified and recognized. Also, the uman ps!cholog! percei"e

s characters b! its o"erall

#. shape and features such as stro$es, cur"es, protrusions, enclosures etc. %hes

e properties, also called

&. Features are e'tracted from the image b! means of spatial pi'el(based calculat

ion. A collection of such

). features, called *ectors, help in defining a character uni+uel!, b! means of an Artificial eural etwor$

-. that uses these Feature *ectors.

Page 3: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 3/17

INTRODUCTION

1.  Automated Optical Character Recognition has gained impetus largel! due to its application in the

2. fields of Computer *ision, ntelligent %e't Recog

nition applications and %e't based decision(

3. ma$ing s!stems. %he approach ta$en to sol"e th

e OCR problem was based on ps!cholog! of the

#. characters as percei"ed b! the humans. %hus the geometrical features of a character and its

&. "ariants were considered for recognition

Page 4: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 4/17

/. Recognition using Correlation Coefficients was basedon the Cross Correlation of input characters or their tr 

ansforms, with the database templates0 so as to acco

mmodate minor differences was used. t introduced F

alse or rroneous Recognition among characters "er ! similar in shape, such as 4 5, 6 4 7, O, 8 4

9 etc.

Page 5: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 5/17

%he solution to this problem lies in A, a s!stem that can percei"e and recognize a character b

ased on its topological features such as shape, s

!mmetr!, closed or open areas, and number of p

i'els.

%he ad"antage of such a s!stem is that it can be

trained on samples and then can be used

to recognize characters ha"ing a similar :not e'a

ct; feature set.

Page 6: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 6/17

 A character can be written in a number of wa!s differing in shape and properties, such as %ilt,stro$e, Cursi"it! and O"erall

shape. A plethora of Fonts are a"ailable for use in an! comm

onl!used <ord =rocessing Application >oftware.

?et, while percei"ing an! te't written in a "ariet! ofwa!s, humans can easil! recognize and read each character.

%his is because the human perception processes the inform

ation b! the features that define a characters shape in an o"

erall fashion.

%hus, while modeling the human perception model in machin

es, a rugged Feature 'traction algorithm is needed before a

n A can be applied for classification of characters

Page 7: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 7/17

RECOGNITION LGORITHM

=re(processing/.  An! image needs some =re(processing, before being fed to the recognition s!st

em. %he first step is the con"ersion of an! $ind of image into a 6inar! image :the

one ha"ing pi'el "alues as 9 4 1 onl!;.%he following flowchart denotes the step

s of the algorithm,

Page 8: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 8/17

Feature 'traction

Feature 'traction ser"es two purposes0 one is to e'tract properties that can identif! a

character 

uni+uel!. >econd is to e'tract properties that can differentiate between similar characte

rs.

 A character can be written in a "ariet! of wa!s, and !et can be easil! recognized correc

tl! b! a

uman.

%hus, there e'ist a set of principles or logics that surpass all "ariation differences. %hus

,

the features used b! the s!stem wor$ upon such properties which are close to the ps!c

holog! of 

the characters.

Page 9: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 9/17

>umcm@ :cBm,p;. :+uation 1;

Figure 2 orizontal Dines at different heights

>imilarl!, a set of "ertical lines drawn at "arious distances along the width, depicting the sum of 

pi'els, can also ser"e as another feature set, as shown. Eathematicall!, the sum of pi'els along

the "ertical line at a width of cBn is gi"en b!,

Page 10: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 10/17

Page 11: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 11/17

 t should be noted that these parameters show the egree of s!mmetr!, i.e. a decimal "alue betw

een 9 :o s!mmetr!; to 1 :=erfect s!mmetr!;, rat

her than %rue or False.

For this, we create a matri', sa! E ha"ing the first

half :horizontal or "ertical; part to be the mirror im

age of the second half.

%hen, the correlation is found between E and .

%his le"el of correlation gi"es us the amount of s!

mmetr! the character has.

Page 12: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 12/17

Figure # orizontal 4 *ertical >!mmetr! in character 

 Another paradigm of character recognition is the number of closed areas in its

shape.

Characters such as A,=, and 8 ha"e one closed area, while others such as 6

and 7 ha"e two.

%here also e'ist characters which are open, such as , -, C etc. %his parameter

also ser"es to broadl! classif! characters based on its openness or closeness.

Page 13: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 13/17

Comparison can be done in !et another form, namel!, using the <% coefficients. A

database was created that contained the <alsh(adamard %ransforms of eachimage template.

%hus, each characters "alue is a"ailable in terms of <% coefficients. 6! calculating

the <% e+ui"alent of the input character, and calculating its correlation to each

sample in the <% database, so maintained, we are able to find the characterha"ing the highest similarit! to the input character.

%hus, classification is done using transformed structure, rather than pi'el(based

arithmetic.

Page 14: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 14/17

  NN Training and Classification

6efore the character recognition can ta$e place, the A is trained, so that it can

de"elop the capabilit! of mapping "arious inputs to the re+uired outputs andeffecti"el! classif! "arious characters.

 For training the A, we use the *ectors generated b! the atabase %emplates

using the abo"e mentioned Feature 'traction techni+ues. %he abo"e mentioned -

different t!pes of features ha"e been used to generate 11 parameters :some of same

t!pe but different "alues;, which are fed to the A.%hus, a matri' of 11'3) "alues is fed to the A to recei"e 3) different "alues at the

output, one for each character in the database.

t ma! be noted that the A uses 6ac$propagation algorithm for Dearning.%he %arget "alues are specified b! the s!stem programmer to accommodate for

small recognition errors, which ma! be changed from application to application.

Page 15: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 15/17

  NN Training and Classification

%he A was trained for 1999 iterations, which too$ around 21 seconds to complete.

%he %raining Function was set to use >um >+uared rror rather than Eean>+uared rror, because the s!stem needed to calculate the effect of Goint errors in

all the parameters, rather than o"erall error.

 An error goal of 9.9991 or 9.91H was achie"ed b! the A.

Figure ) nput(Output parameters of A during %raining

Page 16: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 16/17

EXPERIMENT L RESULTS SUMM R!

%he A s!stem, being trained using standard templates of the capital alphabets

and numbers :as the! are;0 shows 199H recognition for the set of data, it was trainedupon.

%able 1 ( Recognition Rate for Intrained Fonts b! the A s!stem.

Page 17: ARTIFICIAL  NEURAL  NETWORK  BASED  OPTICAL.pptx

8/17/2019 ARTIFICIAL NEURAL NETWORK BASED OPTICAL.pptx

http://slidepdf.com/reader/full/artificial-neural-network-based-opticalpptx 17/17

CONCLUSION

%he A based s!stem has shown promising results due to the fact that despite

being trained onl! on a single set of templates :independent of an! pre(defined font;,it not onl! gets trained in 21 seconds, but also can recognize the fonts :for which it

was not trained; with high efficienc!, as alread! obser"ed.

%he s!stem has its ad"antages such as Dess %ime Comple'it!, *er! >mall atabaseand igh Adaptabilit! to untrained inputs, with onl! a small number of features to

calculate as compared to the method followed in J3K. ?et, the s!stem has a large

scope for further de"elopments.