Preattentive reading and selective attention for document image analysis

Abstract

PixED (from Pixel to Electronic Document) is aimed at converting document images into structured electronic documents which can be read by a machine for information retrieval. The approach is based on the combination of perception and symbol reading which are the two processes involved when humans detect the organisation of a document. "Preattentive reading" denotes the physical segmentation related to perceptual organisation. "Selective attention" means that symbol reading is limited to specific sequences of symbols or to pre-attentively selected locations. An OCR provides the primary structured description of the document. PixED improves the quality of this description, completes the physical segmentation and adds a logical description. A distributed software architecture and an incremental strategy are defined to enable the integration of perception and symbol reading. The approach is tested on a set of documents composed of several pages which are gathered from proceedings of scientific conferences.