|
To view the full report, please click here.
Thesis Abstract
A method
for pattern recognition (character recognition), invariant under translation, rotation and scaling, using Artificial Neural
Network (ANN) classifier is presented here. That is, the system is intended to recognize translated, rotated, scaled or composite
versions of the exemplar patterns (character). The method is a combination of several novel algorithms that decomposes the
task of classification into several steps. The classification process is performed upon a set of thinned characters. The method
in this system consists of two parts. The first is a preprocessor and the second is an ANN classifier. The first step of the
method takes into account the invariant properties of the object and extracts topological object characteristics. This preprocessing
phase uses a robust feature set generated by radial coding and differential radial coding schemes. For an ideal case, rotated
or scaled pattern is assumed to be noise free and distortion less. Hence, the outputs of the preprocessor are invariant to
rotation, scaling and translation. However, in practical case rotation and scaling always insert some noise in the input pattern.
The second part of the method is an ANN classifier, which is trained by the backpropagation algorithm. Outputs of the preprocessor
are fed into this classifier. With slight variation, the classifier is expected to classify the input character patterns correctly.
The uniqueness of this approach resulted from the application of the recognition algorithm on thinned characters. The method
is tested to classify English characters. Performance analysis of the model is presented which shows satisfactory accuracy.
The performance also depends on the number of neurons in the hidden layer. So, the effect of the structure of the Neural Network
classifier on the performance is also studied.
To view the full report, please click here.
|
