The extraction and recognition of text from multimedia document images

TLDR: It is shown that highly accurate and fast recognition can be achieved using a remarkably small number of carefully chosen features, and that after training on only seven quite similar fonts, the recognition algorithm provides greater than 95% accuracy on fonts different to the training set.

Abstract: Almost all the current commercial OCR machines employ matrix matching, resulting in high speed and accuracy, but a severely restrictive range of recognized fonts. Published algorithms conversely, concentrate on feature extraction for font independence, yet they have previously been too slow for commercial use. Current algorithms also fail to distinguish between text and non-text images. This thesis presents a new approach to the automatic extraction of text from multimedia printed documents. An edge detection algorithm, which is capable of extracting the outlines of text from a grey level image, is used to obtain a high level of discrimination between text and non-text. An additional benefit is that text of any colour can be read from almost any background, provided that the contrast is reasonable. The outlines are approximated by polygons using a fast two-stage algorithm. A feature extraction approach to font independent character recognition is described, which uses these outline polygons. It is shown that highly accurate and fast recognition can be achieved using a remarkably small number of carefully chosen features. The results show that after training on only seven quite similar fonts, the recognition algorithm provides greater than 95% accuracy on fonts different to the training set. A more complex edge extraction algorithm is also described. This is capable of extracting text and line graphics from an arbitrary page. Although not essential for character recognition, this algorithm is useful for the interpretation of engineering drawings. As a further contribution to this problem, a thinning algorithm is defined, which is non-iterative and uses the polygonal approximated outlines from the edge extractor.