Overview

Overview
Classification of zones into various syntactic categories such as text, graphics, logo, etc. is an important subtask performed by any generic OCR system. Automated techniques for training zone classifiers are crucial because a) the test datasets keep changing and automated algorithms can be easily adapted to the new datasets by just retraining the algorithms, b) the algorithm is not governed by subjective bias of an individual, c) these methods are quite generic and can be employed for any classification problem.

A decision tree based classifier has been implemented and tested on the UW dataset. The classifier has 96% accuracy and has approximately 33% fewer misclassification errors than the University of Washington algorithm.

• Feature Extraction: Software has been written for extracting features based on connected components. These features include mean and standard deviation of component height, width, area and aspect ratio; number of connected components; percentage of area covered by connected components.
• Classifier: A CART-based decision tree was trained on the University of Washington Dataset.
• Evaluation: The training and testing was done by dividing the dataset into 10 mutually exclusive subsets and training on 9 and testing on 1, and then rotating the test and training sets.