Abstract & Fullpaper-Abet Tool for Visually Impaired Using Artificial Neural Network

8/13/2019 Abstract & Fullpaper-Abet Tool for Visually Impaired Using Artificial Neural Network

http://slidepdf.com/reader/full/abstract-fullpaper-abet-tool-for-visually-impaired-using-artificial-neural 1/6

ABET TOOL FOR VISUALLY

IMPAIRED USING ARTIFICIAL

NEURAL NETWORK

PREPARED BY,

K.J.DHEEPTHI,*IT[FINAL

YEAR],

M.HARINI’,IT[FINAL YEAR],

JEPPIAAR ENGINEERING

COLLEGE,

CHENNAI-!.

EMAIL

ID"#$%%&'$(&)+$/(0.12/

$)(3(('4!/(0.12/

CONTACT NO"*

!5678565,*!!9:495;

ABSTRACT"

Digital Signal Processing

technique is playing a vital role in

S&%%1$ R%123('(23 which is a

<1(3'(3 <(%0# of computer science

& mathematics. The systems designed

using DSP can offer facilities to extract

speech features but doesn’t help in

appropriate decision maing.Thus we

suggest adding an ARTIFICIAL

NEURAL NETWORK.

!isual impairments affect a

large percentage of population in

various ways including 1202)-

#%<(1(%31=, &)%>=2&(, 02? @((23,

and other more severe visual

disabilities."urrent estimates suggest

that there are approximately 64

/(00(23 >0(3# 2) @(00= (/&()%#

(3#(@(#0 (3 INDIA 023%.

T$ 2) &&%) 123(' 2<

$2? A)'(<(1(0 N%)0 N%'?2)+ &0=

@%)= (/&2)'3' )20% (3 /+(3 '$%

VISUALLY IMPAIRED to

recogni#e the pictures by E#%

#%'%1'(23 3# /3= /2)% 1231%&'

where we are including 2) 2?3

innovative concept 2< ##(3 00 '$%%

'%1$3(% (3 HEAD MOUNTED

DEVICE i.e. HELMET (3 2)#%) '2

@2(# VULNERABLE

ACCIDENTS.



FULL VIEW OF THE PAPER:

“An idea tat i! deve"oped and p#t

into action i! $o%e i$po%tant tan an

idea tat e&i!t! on"' a! an idea.()A"*e%t

Ein!tein

He%e a%e !o$e pat)*%ea+in, !o"#tion!

to %ea" "i-e p%o*"e$! ic e ant to

!a%e it te o%"d.

This paper describes a new

approach to image enhancement for people

with severe visual impairments to enable

mobility in an urban environment.$

neural%networ classifier is used to

identify obects in a scene so that image

content specifically important for mobility

may be made more visible.

K%=?2)#" 'ow !ision( )obility

$ids( *ead%mounted Display( *ead%

mounted "amera+*,'),T-( $rtificial

eural etwor.

.BACKGROUND"

• /ver the last fifteen years an

increasing amount of research has

attempted to apply techniques from

computer vision to the needs of

people with low vision.

• Peli and co%worers have been

woring on improving vision in

cases of loss of contrast sensitivity

by contrast enhancement in

specific spatial frequency bands

important to tass such as face

recognition.

• The development of head%mounted

display units has led to some

research in application of these for

use by people with low vision

(a) Original image, (b)

Enhanced by Peli method

. CONTENT DRIVEN

IMAGE ENHANCEMENT"



• 0n 1ristol enhanced images( each

obect in the scene is coloured

using a solid high saturation colour

corresponding to the type of obect(

for example the road is blac(

pavement white(vehicles bright

yellow and so on.

• 0n current wor ( a scheme is used

whereby all obects are classified

as one of a set of eight obect

classes( so for example the

2obstacle3 class contains such

things as bounding walls( lamp%

posts( pillarboxes etc.

(a) Original image, (b) Enhanced by

Bristol method

(a) Blurred image, (b) Peli method, (c)

Bristol method

$rguably the Peli method has

reduced the visibility of the scene by

reduction in the overall contrast of the

image( whereas in the image enhanced by

1ristol method one can clearly see the

boundary between pavement and road(

which is invisible in the other two images.

.8 IMAGE ENHANCEMENT

SYSTEM"

• 0nput to the system would come

from head%mounted camera(which

we implement it in *,'),T and

output will be displayed on a head%

mounted display in such a way that

even the old or low vision people

can travel without any guidance.

• The significance of implementing

eural etwor concepts in

*elmet will mae us to recogni#e

the obects even in weather

conditions lie 4/5()0ST(6

• The image segmentation stage

segments an image into a number

of regions which are deemed to

correspond to a single object or

obect part.

HEAD MOUNTED CAMERA



HEAD MOUNTED DISPLAY

The feature extraction stage takes the

pixels of a region and calculates a set of

numeric features or feature vector3(

which describes the visual properties of

the region and its context in the image.

• The neural net classifier takes the

feature vector of a region as input

and its output corresponds to an

obect class to which the region

belongs( for example road( vehicle

etc.

• The final image labeller stage

produces the coloured image in

which the pixels of each region are

coloured with the high saturation

colour corresponding to the

determined obect class for the

region.

.7 SEGMENTATION"

• This figure shows the quality of

segmentation obtainable

using this 1ristol technique.

(a) !deal "egmentation, (b)

#uto matic "egmentation

• 0ndividual regions are then

extracted using a connected%

components algorithm such that all

connected pixels assigned to the

same cluster belong to the same

region.

.9 FEATURE ETRACTION"

• 0n order to classify regions into an

obect class using a neural networ

classifier( the raw image

information in the form of pixels of

a region must be reduced to a small

set of numeric features which



describe the region sufficiently

well for it to be classified.

• This is achieved by the feature

extraction stage.

.; SHAPE"

• Shape should intuitively be a very

powerful feature for discriminating

obect classes.

•

*ere shape representation is usedwhich is invariant to translation(

rotation( and scaling.

• The representation used here is a

4ourier shape descriptor( which

considers the outer boundary of a

region as a periodic sequence of

two dimensional points( where

transformation to the frequency

domain allows the significant

properties of the shape to be

characterised in a small number of

values.

8.NEURAL NETWORK

CLASSIFIER"

• $ multi%layer perceptron with one

hidden layer is used as a classifier(

with a 78%n%9 architecture(and with

hidden and output units having

logistic sigmoid activation

functions.

7. COLOUR CONSISTENCY"

:hen using colour to recognise

obects( we should be careful to note that

the colour of pixels in an image do not

correspond to the actual colour of an

obect( but it’s perceived colour under a

general unnown illuminant( for

example:#=0($'.

(a)$lassification,(b%ncorrected

$lassification,(c) $orrected $lassification

9. PROCESSING SPEED"

$ut &egion "ilhouette, Boundary

reconstruction from (b) ' coefficients, (c)

coefficients

&egion silhouette, Boundary reconstruction from

(b) ' coefficients, (c) coefficients



• $s /pthamologists are aiming

eventually at a real%time

implementation of our system(

processing speed is important.

• They are able to achieve a

classification accuracy of up to

46.; by area in a processing time

of ust 7;;ms.

;. CONCLUSION"

Thus the co$p#tin, o%"d has a

lot to gain from neural networs..N%)0

N%'?2)+ also contribute to other areas of

research such as 3%)202=

&=1$202=.They are regularly used to

/2#%0 &)' 2< 0(@(3 2)3(/ & to

investigate the (3'%)30 /%1$3(/ of

the >)(3.

Thus our paper proves that how

A)'(<(1(0 N%)0 N%'?2)+ C0(<(%r

plays a very important role in maing the

VI/UALL0 IMPAIRED PEOPLE to

recogni#e the <eal time obects in a road

by ,dge detection(4eature ,xtraction and

many more concepts(where we are

including our innovative concept of adding

all this techniques in a HEAD

MOUNTED DEVICE i.e. HELMET

which avoids the roadside ACCIDENTS

even during )/0ST "'0)$T,.

1It2! o#% *%ain on o#% pape%(3

6.CONTRIBUTION"

*ave discussed our paper in

SHANKAR NETHRALAYA EYE

HOSPITAL,CHENNAI where we have

got an idea that our innovative concept can

also be implementable.

5.REFERENCE"

[]E P%0( !!-( 2I/%

%3$31%/%3' <2) '$% V(00= I/&()%#

S(/0'(23 3# E&%)(/%3'0

R%0', !n*estigati*e

[8]E P%0( !!;-( H%#-/23'%#

#(&0= 02? @((23 (#, %npublished

manuscript .

Documents

Abstract & Fullpaper-Abet Tool for Visually Impaired Using Artificial Neural Network