Data Fusion by Intelligent Classifier Combination for Tamil Isolated Handwritten Characters
No Thumbnail Available
Date
2011-07
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Image classification is an area in image processing where tlie primary goal
is to separate a set of images according to their visual content into one of a number
of predefined categories. It is a vital stage in Tamil Character Recognition System
and in the present scenario any of the many single classification system is being
used.
In order to improve the accuracy of the character classification, the present
research work proposes the use of mulfiple classifier or fusion of classifiers.
Fusion-based classification is a technique that has been proven to be efficient than
single classification algorithms. Fusion of classifiers has several advantages over
single classifiers. It improves the accuracy of classification and reduces the failure
to recognize rates.
The main aim of this research work is to classify Tamil character images
using ftision or multiple classification algorithms. Image classification is defined
as a process which groups similar images together during training and maps an
incoming image whose features matches close to a classified group during testing
stage.
In the present research, a classifier combination, that uses three classifiers,
namely. Neural Network (NN), Support Vector Machines (SVM) and K Nearest
Neighbor (KNN) is proposed. Using these three classifiers, three two-classifier
combinafion (NN + SVM, NN + KNN, SVM + KNN) and one three-classifier
combination are proposed (KNN + SVM + NN).
Six image features, namely, mean, standard deviation, median, area,
minimum and maximum intensity of the image were calculated for each image.
The calculated features were stored as feature vector matrix and was used as input
the classifiers. The outputs from each classifier were combined using a majority
voting scheme to achieve a final decision.
Three performance metrics were used during experimentation. They are
error rate, accuracy and speed of classification. N-fold cross-validation method is
used, where N is varied from 1 to 5 in steps of 1. When N > 1, the average value is
calculation and is taken as the final performance result.
Experiments with various parameters prove that the aggregation of
classifiers to classify images improve the classification results in terms of
accuracy, error rate and speed of classification.
From the results, it is evident that the performance of classifier that
combines KNN with SVM and KNN with NN is better when compared to other
single classifiers and three classifier combinations. While comparing KNN
combined with SVM and NN, the KNN + SVM combination produced better
results than KNN + NN.
While considering the number of runs used, it is evident that the number of
runs and error rate are indirectly proportional to each other. Thus, after comparing
the result of the fusion classification models with their single classifiers, it is clear
that the KNN + SVM classifier shows significant improvement both in terms of
error and accuracy.