You are on page 1of 30

NEURAL NETWORK BASED ENGLISH SCRIPT RECOGNITION SYSTEM

UNDER THE GUIDANCE OF PROF Mrs.V.S.KASHID BY DIVYA GOUD RUCHA KULKARNI SHRUTI SAKHUJA

Contents of Presentation:

Review of related work done. Literature survey. Problem Statement. What are neural networks? Block diagram and description. Kohonens network. Thank you.

Review of Related Work Done!

Handwriting Recognition has an active community of academics studying it. The biggest conferences for handwriting recognition are the International Workshop for Handwriting Recognition (IWFHR), held in evennumbered years, and the International Conference for Document Analysis and Retrieval (ICDAR), held in oddnumbered years. Both of these conferences are organized under the auspices of the IEEE. Active areas of research include: Online Recognition Offline Recognition Signature Verification Postal-Address Interpretation Bank-Check Processing

Continued
ANNs have been extensively used for document analysis and recognition . Most efforts have been taken to recognize isolated handwritten characters. Six different scripts are popularly used all over the world. Of which Roman Script is the most widely used .Roman script is used by languages like English ,German,and French. Handwriting of an individual is unique. And since a neural network learns by example the example must be selected carefully .Script may sometimes be very poorly written.This requires a wide database. We are hence going to create a wide database for the neural network so that it can recognize any form of the roman English script.

Literature Survey

David R Wilson, Applying new technologies to foreign recognition difficulties, PP, Equal Opportunities Dept, Harton School, South Shields. Sachin Ratikant Gengaje, Synopsis Thesis, Dept of Electronic Engineering, Walchand College of Engg, Sangli. Bernhard Wegmann and Christoph Zetzsche, Feature Specific Vector Quantisation of Images. IEEE Transactions on Image Processing, Vol 5, no. 2. A new interband vector quantisation of a human vision based image representation is presented.

Continued

Andrew W Senior and Anthony J Robinson, Members IEEE, An Handwriting Recognition System, PP. IEEE Transactions on Pattern anaylsis and machine intelligence,Volume 20, no.3, March1998. This paper describes a complete system for the recognition of offline handwriting, preprocessing techniques are described, feature extraction and recurrent neural networks.

PROBLEM STATEMENT
In the proposed system we are going to develop the system for automatic identification of handwritten roman script. The system is an offline recognition system. We make use of the Kohonens Self-Organizing Map Algorithm, which is also called as Unsupervised algorithm. This network consists of two layers first is the input layer and second is the output layer. Steps For Self-organizing Map Present the Input Patterns to the network Train the network for the maximum database Adjust the weights after training Decide the number of input and output layer and learning rate as well As number of iterations after training. See the winner index and calculate the accuracy of the network

BLOCK DIAGRAM
Handwri tten script Storing the images Segment ation Fitting of the images Normaliz ation Feature extractio n

Handwri tten script recognis ed

BLOCK DIAGRAM DESCRIPTION


1.

2.

3.

4.

5.

The hand written characters or script is stored in the computer to create a database for training. The single word image is then preprocessed using matlab functions for fitting of the image into a minimun fitting rectangle. In this process some of the variations in the images are removed which do not affect the identity of the word. The image is then converted in a binary form and it is sent to the feature exraction block. In the feature extraction block feature extraction is done using matlab functions.each character has specific features. After feature extraction the image is sent to the recognition block where it is recognised using kohonens algorithm.

Input layer

This is how the scanned image will look. This image is then preprocessed which includes edge detection , dilation functions.

MINIMUM FITTING

After the preprocessing stage the image is fitted into a minimum fitting rectangle. The minimum fitting rectangle is of size 5x7. We make use of a 5x7 block since the image gets spread out in an 8x8 block. A minimun fitting rectangle is used to get a uniform image size and to reduce the number of pixels required to store the image.

FEATURE EXTRACTION

Feature extraction of a character depends on the specific properties of a character.

Each character has different features.


Some specific features of a character are height ,width, stroke etc. A slight change in the feature of a character like stroke can change the definition of the character.

Feature extraction example:

What is a neural network?


Human brain is a group of large number of cells called neurons that process the information. The basic cell is called neuron which acts like a processor.

Continued.
1. 2. 3.

Each neuron consists of: Cell Body also known as Soma Dendrites for incoming information Axon for outgoing information

BIOLOGICAL AND ARTIFICIAL NEURON

TRAINING
How do we train a neural network? The task is similar to teaching a student. How do we do that?

First, we show him some examples.


After that, we ask him to solve some problems.

Finally, we correct him, and start the whole process again.


Hopefully, hell get it right after a couple of rounds

How to train the neural network?


Just like the human brain learns by example i.e it needs to be trained to recognize something. The neural network must be trained too. There are 2 ways the neural network learns: Supervised learning Unsupervised learning We are going to use the unsupervised learning algorithm. This is also known as self organizing feature maps (SOFMs) or KOHONENS NETWORK .

Artificial Neurons

The neuron is a basically a simple calculator.

It calculates a weighted sum of the inputs and applies an activation function to the result.

NEURON TRAINING

Training of neuron requires two parameters the input vector and the weight matrix. The input vector is the vector matrix which we will get after feature extraction. The weight matrix assigned some weights .this assignment is random . Now the minimum euclidean distance is calculated using the formula
D(j)=( ( wij - xi)2 )

Continued

The winning neuron or node is the one which has the minimum euclidean distance. After calculating the winning node the neighours of the node are allowed to adjust their weights. The weights are adjusted using the formula:
Wij(new) = wij(old) + ( xi wij(old) )

Cont
Where , Xi = input vector Wij = weight matrix value D(j) = euclidean distance a = learning rate

The learning rate has to be set first and it will go on reducing as the number of iterations increase. The training performance depends on the number of iterations performed by the network.

KOHONEN NETWORK

MODEL OF A KOHONEN NETWORK


S>Ti

S:
S<Ti

Weighted Sum Threshold Active Neuron (S>Ti)

Ti:

S<Ti

Inactive Neuron (S>Ti)


S>Ti

Normalized Layer

Continued

The kohonens network makes use of the competitive learning rule In The competitive learning rule there is only one neuron which fires for a particular input . Ie the neuron that wins the competition is the output neuron. There can be only one neuron which fires for a particular input vector.

Kohonens network

Number of layers:2 Input nodes:35 Output nodes:more than 26

Learning Vector Quantization

The output units are positioned by adjusting their weights through supervised training to approximate their decision surfaces. The task of forming clusters of input vectors in order to compress the amount of data without losing important information.

FUTURE SCOPE:

Non-cursive script recognition when combined with cursive script recognition can produce a wide database for handwriting recognition. This will not only improve the accuracy of our project but also its applications. Alphabets in the upper case for English language have not been included in our project if include will increase its applications and its value. By increasing the database for training the neural network we can increase the accuracy of the network. Unsupervised training is also used for online character recognition . The same database can be used for online character recognition. Neural networks do not perform miracles. But if used sensibly they can produce some amazing results.

References
REFERENCES FOR PATTERN RECOGNITION: http://www.icdar2007.org http://kopernik-eth.npac.syr.edu:1200/Task4/pattern.html http://www.springerlink.com/content/2y4w3n1emqbcgyta

REFERENCES

http://www.scholarpedia.org/article/Kohonen_network#Introduction#I ntroduction http://www.scholarpedia.org/article/Kohonen_network#Introduction#I ntroduction http://www.willamette.edu/~gorr/classes/cs449/Unsupervised/SOM.ht ml http://www.heatonresearch.com/book/programming-neural-networksjava.html http://www.isical.ac.in/~cvpr/ http://www.generation5.org/content/1999/selforganize.asp http://edman.de/BPR/chapter2.html http://www.borgelt.net/doc/lvqd/lvqd.html#Introduction#Introduction

Our sincere thanks to our guide Prof.Mrs. V.S.Kashid

Thank you.

You might also like