Mini Project, Computer Science Department, College of Engineering Chengannur 2003-2007, Affiliated to Cochin University of Science and Technology (CUSAT), Kerala, India

1. CHARACTER RECOGNITION USING
ARTIFICIAL NEURAL NETWORKS
EVA NINAN
DHANYA S. KURUP
JAISON ABEY SABU
TRESSA JOSE

2. AIM
 To create an ADALINE neural network
 Specific Application – Recognize trained
characters in a given matrix grid
 Develop object oriented programming skill

3. BIOLOGICAL NEURAL NETWORKS
Dendrites
Synapse
Axon

4. ARTIFICIAL NEURAL NETWORKS
 An information-processing system that has certain performance
characteristics in common with biological neural networks.
 Information processing occurs at many simple elements called
neurons.
 Each connection link has an associated weight, which, in an
ANN multiplies the signal transmitted.
 Each neuron applies an activation function (usually nonlinear)
to its net input (sum of weighted signals) to determine its output
signal.

5. A Multi-Layer ANN

6. ARTIFICIAL NEURON
∑
f
a y
x1
x
x
2
N
w
w
w
1
2
N
a w xi i
i
N
=
=
∑1

7. SOME COMMON ANN MODELS
 McCulloch-Pitts Model
 Perceptron
 ADALINE (Adaptive Linear Neuron)
 MADALINE (Many ADALINE)

8. THE ADALINE
The ADALINE (Adaptive Linear Neuron) [ Widrow & Hoff,
1960] typically uses bipolar (1 or -1) activations for its input
signals and its target output. The weights on the
connections from the input unit to the ADALINE are
adjusted. The ADALINE also has a bias, which acts like an
adjustable weight on a connection from a unit whose
activation is always 1.
In general, an ADALINE can be trained using the delta rule
also known as Least Mean Squares (LMS) or Widrow-Hoff
Rule.

9. THE ADALINE - Architecture
Architecture of an ADALINE

10. ADALINE - Algorithm
Step 0: Initialize weights
Set Learning rate
Step 1: While Stopping condition is false,
do steps 2-6
Step 2: For each bipolar training pair s:t,
do steps 3-5
Step3: Set activations of input units, i=1,…..,n;
Xi=Si
Step 4: Compute net input to output unit;
y_in = b + sigma i Xi Wi
Step 5: Update weights, i=1,….,n ;
Wi (new) = Wi (old) + alpha (t-y_in)Xi
Step 6: Test for stopping Condition;
If the largest weight change that occurred in step 2 is smaller than
a specified tolerance, then stop; Otherwise continue

11. FLOWCHART FOR DELTA TRAINING
ALGORITHM

12. REQUIREMENTS
 A basic Operating System (MS DOS or
above)
 Turbo C++

13. DESIGN AND IMPLEMENTATION
The design of the neural network we call
“Neurotron v1.0” involves five stages.
 Implementing the structure
 Training the Artificial Neural Network
 Getting the input to the network
 Processing the data using the ADALINE
Network
 Displaying the output.

14. Implementing the structure
– A single layer, feed forward, fully connected
network is designed and implemented using
neuron and network objects.
– It contains 72 (9x8) input neurons and a bias term
– It contains 8 output neurons to represent the
ASCII code of the recognized alphabet in binary.
– It contains a total of 73x8=584 connections and
weights.

15. Training the ANN
 The ANN is trained using the Delta Rule
mentioned earlier.
 The initial weights are random numbers
between -0.5 and +0.5
 It is currently trained for 70 characters
including 58 ’A’s and one set of ‘B’ to ‘L’.
 The input is given in a 9x8 matrix of 1’s and
0’s.

16. Sample input matrix
{
{0,0,0,1,1,0,0,0},
{0,0,1,0,0,1,0,0},
{0,1,0,0,0,1,1,1},
{1,0,0,0,0,0,0,1},
{1,1,1,1,1,1,1,1},
{1,0,0,0,0,0,0,1},
{1,0,0,0,0,0,0,1},
{1,0,0,0,0,0,0,1},
{1,0,0,0,0,0,0,1},
},

17. GETTING THE INPUT TO THE
NETWORK
 Input is received on a Black and white grid by
mouse clicks

18. PROCESSING THE DATA
 The Neurotron v1.0 loads the input,
propagates the network and calculates and
displays the output on clicking the generate
button in the GUI.

19. 5 DISPLAYING THE OUTPUT
 The output is displayed in the Recognized
Character Box given on screen.

20. FLOWCHART OF THE PROGRAM

21. SCREENSHOT OF THE PROGRAM

22. RESULTS AND FUTURE SCOPE
 The Neurotron v1.0 is currently trained to identify 70
characters consisting of 58 ‘A’s and one set of
characters ‘B’ to ‘L’.
 May be further trained to recognize any other
character set by training it with a suitable character
set.
 Learning capability is limited by the number of
neurons and connections in the system. Training
with very large character sets may result in the
weights not converging i.e. the net may be unable to
learn the entire set.

23. FURTHER IMPROVEMENTS
 The network can be trained for a wide range of other
characters, using optimal training set.
 The number of input and output layers may be increased as, in
the current system, weights may not converge during large
training sets. This can be done by changing the way getting the
output. Instead of getting the ASCII of the character, the output
may be only one neuron with output ‘1’ for each character.
 Another way of increasing the power o the neural network is to
add one or more hidden layers to the network and use the back
propagation algorithms and training them using the Back
propagation training algorithm.
 The application and trainer can be integrated to form a
complete flexible software.

24. APPLICATIONS
 Language Processing
 Image and Audio Processing
 Finance and Marketing
 Control Systems
 Database
 Weather forecasting
 Other

25. THANK YOU