Khulna University ofEngineering & Technology
Department of Electrical and Electronic Engineering
Seminar On EE-4130
Presented by–
S.M.Kamrul Hasan
Roll No. 1003079
Department of
Electrical & Electronic Engineering 1
American Sign Languagewordrecognitionwith
a sensoryglove usingartificial neural networks
Published in: Elsevier Journal of Engineering Application of
Artificial Intelligence, Volume 24, Issue 7.
Date of Publication: October 2011
Pages: 1204-1213
Authors:
Cemil Oz & Ming C.
Leu
2
Outline :-
 Problem definition
 Motivation
 What is Sign Language
 Why Data glove
 System Structure
 Data Collection
 Feature Extraction
 Artificial Neural Network(ANN)
 Training Algorithm
 Test Results
 Conclusion
3
Problem definition:
 Humans have been endowed by nature with the voice
& hearing capability...
Kids speaking
4 /40
Problem definition:-
4
Problem definition:
 Humans have been endowed by nature with the voice
& hearing capability...
 ...but not everybody possesses this capability←deaf people
Deaf kid
4 /40
Problem definition:-
5
Motivation:-
A Real time Sign Language Recognition System
Combines Adaptive Filtering & Artificial Neural Network
 Interpret Sign language into English word.
 Fully Flexible
 Intelligent online learning
 Training time so faster
 Better performance
 Adaptable
6
What is Sign Language ?
 Visual gestural communicating
language used by deaf.
 Movement not just with hands
 Varies from country to country,
language to language
English alphabet in Sign Language
7
System Structure:-
8
System Architecture
Why Data Glove?
 Cyber-Glove
 measure hand shape accurately
 light in weight
 high resolution data
 data acquiring is more difficult
 complicated data processing
 slower recognition rate
Vision based system
 Flock of Birds(3D Motion tracker)
 tracks hand orientation & position
Cyber glove
1.Data Collection:
10
Data collection Block
 Indicate hand moving or static
Cyber-glove
& motion
Tracker
Velocity
Network
X
Y
Z
Data
store
Velocity Network
Velocity
2.Feature Extraction:-
 to determine exactly which features are Important
 Part of the data reduction process
11
 Seven feature vectors
Artificial Neural Network (ANN):
 A Genetic Algorithm, resembles human brain
12
 acquires knowledge through learning.
 it involves human like thinking.
 they handle noisy or missing data.
 BackPropagation
Training Algorithm
 Successful approach to construct ANN
 A supervised learning
Predicted output != actual output
Weight is adjusted until no error
13
Error Adjust
N.N
Compare
Actual
output
Desired output
Input
output
Weight
14
Test Results
 50 different American signs each with 6 samples total of 50x6=300
samples for training.
 Successfully recognize sign language to English Word
Sign Recognition by system
Test Results
15
ANN Test results for Known words
16
Output
Output
Decoding
Does it cross
threshold?
Yes
Training
by N.N
No
Unknown
word
Do you want to
add the word?
Yes
Unknown Words Recognition
Test Results
Test Results
Levenberg-Marquardt Vs. Backpropagation
 Backpropagation gives Better performance with
 less train time
17
Number of hidden layer nodes
18
Trained by 3 users Trained by 7 users
Algorithm Same users New users Same users New users
LM 87.96% 72.22% 92.09% 82.17%
BP 93.51% 80.09% 95.72% 85.41%
Levenberg-Marquardt Vs. Backpropagation
 Better accuracy
 Backpropagation gives Better performance with
Test Results
Conclusion
 An advanced initiative to recognize American Sign Language with
faster training, better accuracy & better recognition performance.
 The ultimate goal of this paper is to further improve the proposed
sign language recognition system that can use sentence recognition
and eliminate the limitations and use it successfully for Human to
Machine Interface for disable people.
19
THANK YOU
ALL
20

Bengali Sign Language

  • 1.
    Khulna University ofEngineering& Technology Department of Electrical and Electronic Engineering Seminar On EE-4130 Presented by– S.M.Kamrul Hasan Roll No. 1003079 Department of Electrical & Electronic Engineering 1
  • 2.
    American Sign Languagewordrecognitionwith asensoryglove usingartificial neural networks Published in: Elsevier Journal of Engineering Application of Artificial Intelligence, Volume 24, Issue 7. Date of Publication: October 2011 Pages: 1204-1213 Authors: Cemil Oz & Ming C. Leu 2
  • 3.
    Outline :-  Problemdefinition  Motivation  What is Sign Language  Why Data glove  System Structure  Data Collection  Feature Extraction  Artificial Neural Network(ANN)  Training Algorithm  Test Results  Conclusion 3
  • 4.
    Problem definition:  Humanshave been endowed by nature with the voice & hearing capability... Kids speaking 4 /40 Problem definition:- 4
  • 5.
    Problem definition:  Humanshave been endowed by nature with the voice & hearing capability...  ...but not everybody possesses this capability←deaf people Deaf kid 4 /40 Problem definition:- 5
  • 6.
    Motivation:- A Real timeSign Language Recognition System Combines Adaptive Filtering & Artificial Neural Network  Interpret Sign language into English word.  Fully Flexible  Intelligent online learning  Training time so faster  Better performance  Adaptable 6
  • 7.
    What is SignLanguage ?  Visual gestural communicating language used by deaf.  Movement not just with hands  Varies from country to country, language to language English alphabet in Sign Language 7
  • 8.
  • 9.
    Why Data Glove? Cyber-Glove  measure hand shape accurately  light in weight  high resolution data  data acquiring is more difficult  complicated data processing  slower recognition rate Vision based system  Flock of Birds(3D Motion tracker)  tracks hand orientation & position Cyber glove
  • 10.
    1.Data Collection: 10 Data collectionBlock  Indicate hand moving or static Cyber-glove & motion Tracker Velocity Network X Y Z Data store Velocity Network Velocity
  • 11.
    2.Feature Extraction:-  todetermine exactly which features are Important  Part of the data reduction process 11  Seven feature vectors
  • 12.
    Artificial Neural Network(ANN):  A Genetic Algorithm, resembles human brain 12  acquires knowledge through learning.  it involves human like thinking.  they handle noisy or missing data.
  • 13.
     BackPropagation Training Algorithm Successful approach to construct ANN  A supervised learning Predicted output != actual output Weight is adjusted until no error 13 Error Adjust N.N Compare Actual output Desired output Input output Weight
  • 14.
    14 Test Results  50different American signs each with 6 samples total of 50x6=300 samples for training.  Successfully recognize sign language to English Word Sign Recognition by system
  • 15.
    Test Results 15 ANN Testresults for Known words
  • 16.
    16 Output Output Decoding Does it cross threshold? Yes Training byN.N No Unknown word Do you want to add the word? Yes Unknown Words Recognition Test Results
  • 17.
    Test Results Levenberg-Marquardt Vs.Backpropagation  Backpropagation gives Better performance with  less train time 17 Number of hidden layer nodes
  • 18.
    18 Trained by 3users Trained by 7 users Algorithm Same users New users Same users New users LM 87.96% 72.22% 92.09% 82.17% BP 93.51% 80.09% 95.72% 85.41% Levenberg-Marquardt Vs. Backpropagation  Better accuracy  Backpropagation gives Better performance with Test Results
  • 19.
    Conclusion  An advancedinitiative to recognize American Sign Language with faster training, better accuracy & better recognition performance.  The ultimate goal of this paper is to further improve the proposed sign language recognition system that can use sentence recognition and eliminate the limitations and use it successfully for Human to Machine Interface for disable people. 19
  • 20.