Feature Extraction and Dimensionality Reduction Models for Printed Numerals Recognition

Abstract

Feature extraction is an important aspect in the field of pattern recognition and machine learning. Processing an image directly by a software tool is not possible due to the size of the image which contains a large amount of data to process. A perfectly designed feature extraction method increases the overall classification process either in machine learning or pattern recognition. Some kind of statistical or geometrical features from images are extracted and are inputted to machine learning algorithms for effective processing. The work presented here uses a novel feature extraction method for dimensionality reduction for classification of Odia numerals. The proposed method converts input images into equivalent matrices; row wise decimal values and produces the feature vector of the input image as a one dimensional array of numerical equivalent of each OCR Odia numeral. The primary attributes of this method is that not only it reduces the dimensionality of the input pattern but also supports image reconstruction. The effectiveness of the method has been experimented for the printed Odia numerals. The recognition rate achieved through the experiment is 81.9%.