Natural Character Recognition Using Image Processing Techniques

Abstract

If we examine our environment we will recognize symbols that we commonly use in both language and numerical systems. Based on this simple observation it has been claimed that there exists a relationship between how often a symbol occurs in nature and how often it occurs in language systems, which we will call the Natural Symbol Hypothesis. Additionally it has been demonstrated that many language and numeric systems use no more than three line segments to represent symbols resulting in 36 different visually unique symbols. To determine the validity of the Natural Symbol Hypothesis,we use image processing techniques, a new classification algorithm based on graph theory and singular value decomposition, to autonomously identify and classify symbols based on geometric and algebraic characteristics. We will demonstrate that although there may be 36 visually unique symbols these symbols can be geometrically characterized into 27 groups consisting of 20 unique symbols and 7 groups made up of the remaining 16 symbols. We will then use edge detection and our classification algorithm to classify over 100 nature scene images from three major categories: National Geographic images, building images, and ancestral images. Finally, we will demonstrate that simpler symbols like ?L?, ?T?, ?X?, and ?Z? are the most commonly identified symbols regardless of what category an image belongs to or what edge detection threshold is used to analyze it.