Semantics-based content extraction in typewritten historical documents

Abstract

This paper presents a flexible approach to extracting content from scanned historical documents using semantic information.

An Efficient Word Segmentation Technique for Historical and Degraded Machine-Printed Documents

“…The determination of word breaks is made in a manner that adapts to the writing style of the individual. For the case of historical machineprinted documents, Antonacopoulos and Karatzas [7] calculate and analyze the horizontal projection profile to identify suitable spaces between words. In the work of Gatos et al [8], word segmentation in historical machine-printed documents is based on a run length smoothing in the horizontal and vertical directions.…”

Flexible Text Recovery from Degraded Typewritten Historical Documents

“…The authors, with other colleagues in their laboratory, have implemented and experimented with various such thresholding techniques [5]. That work demonstrated that the indiscriminate application of any thresholding approach (global or local) does not yield as good results as when a method is applied only to the segmented text.…”

Automatic unsupervised parameter selection for character segmentation

“…Features based upon a convex hull are insensitive to character fonts and sizes, the touching-character problem of various fonts and sizes can be handled even for heavily touching characters or italic-type overlapping characters without slant correction. Table 1 summarizes the characteristics of those approaches [1,8,10,14,4] mentioned above.…”

“…The most known of these segmentation algorithms are the following: projection analysis, connected component analysis, Run Length Smoothing Algorithm (RLSA), contour shape analysis and Hough transform. Representative examples of character segmentation methodologies are the following: Antonacopoulos and Karatzas [1] use the horizontal projection profile of each word segment for character segmentation in historical machine-printed documents. This approach cannot handle the case of overlapping characters.…”

“…The parameters are either user-specified and no training method is included [1,8,10,14,4] or selected through a training procedure over a set of "optimal" parameter values that are usually manually selected based on some assumption regarding the training data [11], [6]. In general, automatic selection of the free parameters is actually an optimization problem [2].…”