Transcript Presentation - Vision and Image Science Lab

Technion - Israel Institute of Technology
Department of Electrical Engineering
The Vision Research and Image Science Laboratory
Handwritten Character Recognition
Using
Artificial Neural Networks
Shimie Atkins & Daniel Marco
Supervisor: Johanan Erez
November 1998
Project Goals
 Creating
an optimal neural network capable of
identifying characters.
 Creating
training files, in which the characters
are reduced, and presented to the net.
 Creating
an application that enables the user to
define forms that are filled and later on identified.
Theoretical Background
What is a neuron?
 A black box that has inputs.
These inputs are summed according to their respective
weights - and the sum is the total input.
 A certain mathematical function is then applied to the
total input, which now becomes the neuron’s output.
What is a Layer?
A group of neurons that are parallel to each other, with no
connections amongst themselves.
Background (cont)
What is a Network?
A group of layers, where neurons’ outputs in one layer
are the inputs of the neurons in the next layer.
The input of the network is the first layer’s input.
The output of the network is the last layer’s output.
Note:
 Our network is fully connected, thus all the neurons in
one layer are connected to each neuron in the next layer.
Training Algorithm
We used the back-propagation algorithm, which
works as follows:
 Net receives input vector and desired output vector.
 Propagate input vector through net.
Compare net’s output vector with desired vector, if they
match within the allowable error --> Done.
Update weights beginning with last layer and advancing
backwards towards the first layer.
Go back to second step.
The Reduction Algorithm
 Divide height and width of the large picture by those of
the small picture.
Each block of pixels received in the large picture
represents one pixel in the small picture.
Count number of “On” pixels in block.
Determine, according to a predefined threshold,
whether the pixel in the small picture is “On” or “Off”.
Note:
 If Dimensions of block are not integral padding is
initiated.
Reduction Algorithm (cont)
Code Implementation
•We Implemented the network in C++.
•We Defined three classes: CNeuron, CLayer, Cnet .
They have the following relationships:
Layer
Neuron
Net
Results of Form Identification
The following form was created (using our
interface) and scanned:
Results (cont)
First net identified the form as follows:
Configuration of Network:
37 nets (8x10 inputs, 3 layers)
Alphabetical - 40 neurons in 2nd layer
Numerical - 25 neurons in 2nd layer
Train File:
dan1to6.trn (40 examples of each
character).
Identification Rate (for check file):
96%
Results (cont)
Second network identified the form as follows:
Configuration of Network:
2 nets (8x10 inputs, 3 layers)
Alphabetical - 30 neurons in 2nd layer
Numerical - 30 neurons in 2nd layer
Train File:
dan1to9.trn (60 examples of each
character)
Identification Rate (for check file):
95.5%
Conclusions
 The identification success rate is proportional to the
number of examples of each character given.
 More than one hidden layer does not necessarily
improve performance.
Most errors are with letters of similar shape (such as IJ or O-D), numerical digits are usually successfully
identified.
The configuration and training parameters of the net
have minimal influence (once inside certain optimal
limits ).