Whole Word and Number Reading Machine Based on Two Dimensional Low Frequency Fourier Transforms.

摘要

The Fourier transform is investigated as a means for developing an optical reader capable of reading a large vocabulary without segmenting the image of a word into individual characters. The reader is capable of reading printed and cursive font styles, is scale invariant, and is substantially insensitive to noise. The image of a particular word is treated as a single symbol; the two dimensional low frequency Fourier coefficients (assuming n coefficients are calculated) define the word's location on an n dimensional hypersphere of unit radius. The distance between individual locations (words) categorizes similar and dissimilar words. The smaller the distance, the more similar images are. Multiple images of a word using various font styles form a unique cluster on the surface of the hypersphere. The distance between clusters (different words) is greater than the distance across a cluster (same word in different font styles). Therefore, by using the centroid of these clusters to build a library of words, input or test words match to the nearest cluster centroid using a minimum distance calculation. This algorithm is capable of correctly recognizing at least 5000 words using 24 various font styles (120,000 individual images). (KR)