A method of document‐image segmentation based on projection profiles, stroke densities and circumscribed rectangles

Abstract

A method is proposed wherein printed documents are segmented into the following three areas: headline area, textline area, and attached area. Then the character lines constituting headline and textline are extracted. This paper describes: 1) the combination of the global features of the document such as projection profiles and stroke densities, and the local features of the document such as circumscribed rectangles are used; 2) the basic features of document elements such as character line periodicity in the textline area, so that a single method can be applied to segment document elements in differently formatted documents. An experimental result of this method applied to eight Japanese and English printed documents of both horizontal and vertical lines showed that 98.4 percent of the headlines in headline area, 99.9 percent of the textlines in textline area, and 83.3 percent of diagrams in the attached area were segmented successfully. The result confirms that this method is effective in segmenting document elements, and applicable to a variety of documents. Part of document processing is implemented in the experiment by special hardware for speed.

<title>Character string detection algorithm using horizontal boundaries, and its application to a part number entry system</title>

“…The conventional methods for detecting character strings are as follows: ( 1) a method using projection histograms (the histograms are calculated several times while changing the range and scanning direction), (5) (2) a method using the properties (e.g. size and ratio of height to width) of components of connected black pixels,(6) (3) a method using the properties of components of connected black pixels in a pre-processed (e.g.…”

A hierarchical method for block segmentation and classification of general document images

“…The preprocessing step shown in Fig. 3 performs noise reduction, correction of an inclined image [4], and so forth. The run length smoothing step decreases the number of connected components in the document image so that the connected component extraction and hierarchical segmentation processing are sped up.…”

“…Over the past few years, several approaches have been proposed: methods which use the structural or statistical properties of document images [l, 2, 5-81, methods which use knowledge processing techniques [ 101, interactive processing methods [3], and a hybrid method which uses both the properties of image pixels and knowledge processing techniques [4]. But problems remain with the generality or robustness of the algorithms.…”

Robust and fast text‐line extraction using local linearity of the text‐line

“…In the method of creating the connected components of circumscribing rectangles or black pixels, it is pointed out that the incorrect extraction might occur if the contact existed between characters or between character and picture [6]. This problem will be overcome by considering the local orientation of text lines.…”