Presentation_OCR

BIRLA INSTITUTE OF TECHNOLOGY
MESRA
(DEOGHAR CAMPUS)
PROJECT TOPIC: CHARACTER RECOGNITION
USING NEURAL NETWORK
AVINASH ABHISHEK (BE/6030/10)
SAMEER V BHALERAO (BE/6082/10)
PULKIT KUMAR (BE/6083/10)
MENTOR: Prof. D.S.ACHARYA

CONTENTS
1. INTRODUCTION
2. PROBLEM DESCRIPTION
3. METHODOLOGY
4. CONCEPT USED IN MATLAB IMPLEMENTATION
5. WORK DONE RESULT
6. DISCUSSION

INTRODUCTION
ARTIFICIAL NEURAL NETWORK(ANN)
AN ARTIFICIAL NEURAL NETWORK IS A
MATHEMATICAL MODEL INSPIRED BY BIOLOGICAL
NEURAL NETWORKS
ANN ARE USED FOR MODELING COMPLEX
RELATIONSHIPS BETWEEN INPUTS AND OUTPUTS

OCR (OPTICAL CHARACTER RECOGNITION)
OPTICAL CHARACTER RECOGNITION (OCR) IS A
TYPE OF DOCUMENT IMAGE ANALYSIS WHERE A
SCANNED DIGITAL IMAGE THAT CONTAINS
EITHER MACHINE PRINTED OR HANDWRITTEN
SCRIPT IS INPUT INTO AN OCR SOFTWARE
ENGINE AND TRANSLATING IT INTO AN
EDITABLE MACHINE READABLE DIGITAL TEXT
FORMAT.

HOW DOES AN OCR WORK??
TWO BASIC METHODS:
1) MATRIX MATCHING: IT COMPARES WHAT THE OCR
SCANNER SEES WITH A LIBRARY OF CHARACTER
MATRICES.
2) FEATURE EXTRACTION: THIS METHOD VARIES BY
HOW MUCH “COMPUTER INTELLIGENCE” IS
APPLIED BY THE MANUFACTURER.

CHARATER SEGMENTATION AND
CHARACTER RECOGNITION ARE
TWO DIFFERENT SYSTEMS
OUR MAIN CONCERN-
CHARACTER RECOGNITION

ELEMENTS AND CLASSIFICATIONN
OF NEURAL NETWORK
THREE LAYERS
 INPUT LAYERS
 HIDDEN LAYERS
 OUTPUT LAYERS
CLASSIFICATION
 FEED FORWARD
 RECURRENT

NEURAL NETWORK APPLICATIONS
AND BENEFITS
APPLICATIONS:
SCANNED CHARACTER RECOGNITION
FACE RECOGNITION
MEDICAL DIAGNOSIS OF BRAIN
BENEFITS:
UNIFORMITY OF ANALYSIS & DESIGN
FAULT TOLERENCE
BUILT-IN CAPABILITY TO ADAPT THEIR SYNAPTIC
WEIGHTS TO CHANGES IN SURROUNDING

2. PROBLEM DESCRIPTION
RECOGNITION OF SCANNED CHARACTER BY
CONVERTING IT INTO MATRIX FORM
NETWORK IS TRAINED BY CREATING A MATRIX
20x20 ELMENTS FOR EACH CHARACTER AND
THEN CONVERTING INTO A COLUMN MATRIX
FORM (400x1)

COST FUNCTION
CONCEPT USED IN MATLAB IMPLEMENTATION
COMPUTES THE COST AND THE GRADIENT OF
THE NEURAL NETWORK
NEURAL NETWORK IS TRAINED WITH THE HELP OF
MINIMISING THE COST FUNCTION USING AN
OPTIMISER CALLED fmincg.
𝐽 𝜃
=
1
𝑚
𝑖=1
𝑚
𝐾=1
𝐾
[−𝑦𝑘
(𝑖)
log((ℎ 𝜃 (𝑥 𝑖
))𝑘) − (1 − 𝑦𝑘
𝑖
)log(1 − (ℎ 𝜃(𝑥 𝑖
))𝑘)]
+
𝜆
2𝑚
[
𝑗=1
25
𝑘=1
400
(𝜃𝑗,𝑘
𝑙
)2
+
𝑗=1
10
𝑘=1
25
(𝜃𝑗,𝑘
2
)2
]

BACK PROPAGATION
ERROR CALCULATION BETWEEN OUTPUT
ACTIVATION AND GIVEN RESULT
PROPAGATES THE ERROR FUNCTION ACROSS THE
HIDDEN LAYERS CORRESPONDING TO THEIR
EFFECTS ON OUTPUT
ONLY FOR FEED-FORWARD NETWORKS
USE OF SIGMOID FUNCTION y=
1
1+𝑒(−𝑥)

3.METHODOLOGY
TRAINING A NEURAL NETWORK
 RANDOMLY INITIALIZE WEIGHTS
 IMPLEMENT FORWARD PROPAGATION TO GET
hΘ(x(i)) (OUTPUT) FOR ANY x(i)
 IMPLEMENT CODE TO COMPUTE COST FUNCTION
J(Θ)
 IMPLEMENT BACKPROPAGATION TO COMPUTE
PARTIAL DERIVATIVES
𝜕
𝜕𝛩jk
(l)J(Θ)

USE GRADIENT CHECKING TO COMPARE
𝜕
𝜕Θ
(l)
jk
J(Θ)
COMPUTED USING BACKPROPAGATION vs.
USING NUMERICAL
ESTIMATE OF GRADIENT OF J(Θ)
THEN DISABLE GRADIENT CHECKING CODE
USE GRADIENT DESCENT OR ADVANCED
OPTIMIZER TO MINIMIZE J(Θ)

TOO FEW HIDDEN LAYERS WILL CAUSE ERRORS IN
ACCURACY ,AND MORE HIDDEN LAYERS INCREASES
COMPLEXITY OF SYSTEM

DISCUSSION
 THE PROJECT HELPS IN EASY AND EFFECTIVE
MODIFICATION OF SCANNED DOCUMENTS
 NEURAL NETWORK IS TRAINED WILTH ALMOST 98%
ACCURACY

FUTURE PROSPECTS
IDENTIFICATION OF SCANNED ALPHABETS BY
TRAINING THE NEURAL NETWORK WITH PRE-
REQUISITE TRAINING SETS
FOR APPROPRIATE VALUES OF EPSILON,THE SYSTEM
WILL BE FURTHER EXPANDED TO INCREASE
ACCURACY OF PROBABILITY DENSITY FUNCTION AND
IMPROVE THE SYSTEM FURTHER.

REFERENCES
1) SIMON HAYKIN-NEURAL NETWORK
2) ADVANCED MACHINE LEARNING –STANFORD
UNIVERSITY
3) INTERNATIONAL JOURNAL OF ADVANCED
RESEARCH IN COMPUTER ENGINEERING &
TECHNOLOGY
VOLUME 1, ISSUE 4, JUNE 2012
4) INTERNATIONAL JOURNAL OF ENGINEERING AND
TECHNOLOGY (IJERT) Vol. 1 Issue 4, June –2012

Presentation_OCR

More Related Content

What's hot (20)

Viewers also liked (16)

Similar to Presentation_OCR (20)

Recently uploaded (20)

Presentation_OCR