A New Dataset Size Reduction Approach for PCA-Based Classification in OCR Application

Abstract

A major problem of pattern recognition systems is due to the large volume of training datasets including duplicate and similar training samples. In order to overcome this problem, some dataset size reduction and also dimensionality reduction techniques have been introduced. The algorithms presently used for dataset size reduction usually remove samples near to the centers of classes or support vector samples between different classes. However, the samples near to a class center include valuable information about the class characteristics and the support vector is important for evaluating system efficiency. This paper reports on the use of Modified Frequency Diagram technique for dataset size reduction. In this new proposed technique, a training dataset is rearranged and then sieved. The sieved training dataset along with automatic feature extraction/selection operation using Principal Component Analysis is used in an OCR application. The experimental results obtained when using the proposed system on one of the biggest handwritten Farsi/Arabic numeral standard OCR datasets, Hoda, show about 97% accuracy in the recognition rate. The recognition speed increased by 2.28 times, while the accuracy decreased only by 0.7%, when a sieved version of the dataset, which is only as half as the size of the initial training dataset, was used.