AN INVESTIGATION OF THE USE OF CONTEXT IN CHARACTER RECOGNITION USING GRAPH SEARCHING,
Abstract
In the paper context is introduced into a character recognition system. The recognition system is considered to be reading a text rather than isolated characters. Two types of context are considered: the syntax, or formal grammar of the text under consideration, and the statistical distribution of the letters in the text. For syntax, a single 'regular' grammar is used. The statistical distribution is approximated by (a) the probabilities of individual letters, (b) the probabilities of letter-pairs (or diagrams), and (c) the probabilities of letter-triplets (or trigrams). For each character input, the character recognizer outputs a list of alternative decisions with their associated 'confidences'. Based on an entire input string, both types of context are then used to make the final decision, using a graph-searching procedure. Experiments were run on a computer with a simulated character recognizer. The results indicate that the graph-searching formulation of the problem does indeed allow the syntax and statistical distribution of the letters to be utilized. The error rate of the character recognition system is reduced when digram and trigram statistics are used; their effectiveness varies inversely as the uncertainty in the statistical distribution of the letters in the text. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 01, 1968
- Accession Number
- AD0707548
Entities
People
- Carl Spencer Christensen
Organizations
- Cornell University