Document image analysis and recognition

Document image analysis and recognition is a scientific discipline that brings together a range of computer techniques aimed at reconstructing the content of a document from its image. While it has long been confined to the field of character recognition, it now has much broader objectives, ranging from simple document classification to complete content interpretation, indexing and re-editing. Thus, the ultimate goal of document image recognition is to generate a high-level representation in the form of structured documents, in a form suitable for the intended application.

By way of introduction, let's consider a page from a scientific book (figure 1 a ) that needs to be "hypertextualized", i.e. produced as an electronic version with hypertext links for navigation. In such an application, it is imperative to determine the logical structure of the book, i.e. its hierarchical organization into chapters, sections and paragraphs, and to identify definitions, exercise statements, experiment descriptions, formulas, etc. Figure 1 b visually reflects this structure at page level, while figure 1 c illustrates the resulting hierarchical structure. It is this structure that can be used for hypertext navigation.

Traditionally, document recognition has been applied primarily to paper documents for which no electronic form was available. Today, these techniques are recognized as being particularly useful for restructuring unstructured or poorly structured electronic documents, using the image produced synthetically, for example with a Postscript print engine.

From a historical point of view, it is interesting to note that optical character recognition predates the development of computer technology, since patents were already filed in the 19th century and a demonstration prototype was reported in 1916. The first computerized approaches to character recognition date back to the early 1960s; for example, the first mail sorting machine (limited to typed addresses) was installed in the USA in 1965. However, major developments date back to the advent of office automation in the 1980s [1] , with the advent of personal computers, graphics screens, laser printers and, above all, flatbed scanners. Since then, practical applications have continued to grow; the considerable increase in information storage capacities and, at the same time, the reduction in their cost has created gigantic needs for the creation of digital libraries, online documentary...