Transcript 投影片 1

Implementing a reliable neuro-classifier
for paper currency using pca algorithm
老師 : 楊士萱 博士
學生 : 陳柏源 碩二
Outline




Introduction
Method
Conclusion
Reference
Introduction(1/1)



Using Principal Components Analysis (PCA) method in order to increase
the reliability of paper currency recognition machines，which use neural
network classifier.
A Learning Vector Quantization (LVQ) neural network model is used as the
main classifier and a total number of 10 US bill types including 1、2、5、
10、20、50 and 100 dollar (new and old model) are considered as
classification categories
The experimental results show a 30% growth in reliability after using
extracted features
Method(1/6)




Preprocessing
Feature extraction
Classification and reliability evaluating
Experimental results
Method(2/6)
preprocessing data

Original image 10x170 array  6x30 array





5 sensor are used，each of them uses two different waves lengths for generating two
channels of data
By using a linear function to generate a new channel of data base on two channel of each
sensor.
Totally 15 channel are obtained among them we select 6 main channel.
A simple algorithm is used to reduce the size of data from 170 pixels in each channel to 30.
Normalization :
Method(3/6)
feature extraction

Using PCA to extract the features of training data

Covariance matrix -> M eigenvector corresponding to the M largest eigenvalues
 Select 30 main features
Method(4/6)
classification


Kohonen’s LVQ is a supervised learning algorithm with the competitive network.
The network has a number of 30 neurons in the input layer and 400 neurons in the
output layer.
Method(5/6)
reliability evaluating


Using specific algorithm to evaluate the reliability of classification
All codebooks are drawn supposing a Gaussian distribution
Method(6/6)
Experimental results


Training data : 3,570 sample data from 40 different class
Testing data : 1,200 sample (30 samples per class)
Conclusion(1/1)


Incrementing the number of codebooks will make the variance of data
within each class and consequently the overlap zone between classes, to be
decreased.
PCA increase the variance within the new components space, but as the
distance between codebooks are increased, it makes the overlap between
pro. densities to be significantly decreased and consequently the reliability
of the system is improved.
Reference(1/1)

Ali Ahmadi,Sigeru Omatu,Michifumi Yoshioka,”Implementing a Reliable
Neuro-Classifier for Paper Currency Using PCA Algorithm,”SICE 2002
Aug.5-7,2002,Osaka.