This thesis is about the evaluation of feature extraction methods for numerals recognition. Three extraction methods are designed and tested. The unconstrained numeral images are imported from the image database. The set of 128 x 128 bits raw images from the database, which consisted large variety digits written by different authors will be used.
The image needed to go through the pre-processing and normalisation phase before they are extracted with different methods. With pre-processing and normalisation process undergo, the image would be easier to recognise.
The three extraction methods are evaluated and compared with the speed and performance of the extraction. The entire image processing procedure including feature extraction are coded and implemented in C. The extracted feature points will be sent to a neural network that capable of recognising numeral character. The Matlab’s neural network toolbox was employed to implement the training part of the back-propagation network. Feature extraction methods will be tested with combination of different preprocessing and normalisation techniques. Testing and modifying the extraction methods will undergo during experiment, so the best feature extraction method will be evaluated.