The T-Recs Table Recognition and Analysis System

Abstract

This paper presents a new approach to table structure recognition as well as to layout analysis. The discussed recognition process differs significantly from existing approaches as it realizes a bottom-up clustering of given word segments, whereas conventional table structure recognizers all rely on the detection of some separators such as delineation or significant white space to analyze a page from the top-down. The following analysis of the recognized layout elements is based on the construction of a tile structure and detects row-and/or column spanning cells as well as sparse tables with a high degree of confidence. The overall system is completely domain independent, optionally neglects textual contents and can thus be applied to arbitrary mixed-mode documents (with or without tables) of any language and even operates on low quality OCR documents (e.g. facsimiles). S . -W. Le e an d Y . N akan o ( E d s. ): DA S ' 98, LNCS 1655, p p . 255-270 , 1999. c S p r i n ge r-Ve rl ag Be rl i n He i d e l b e rg 1999 Thomas Kieninger and Andreas DengelWords are our elementary objects. Their bounding box geometry and optionally their textual contents constitute the input of the system. Formally, a word is described as a triple W = (T, G, A), where T keeps the textual contents, G denotes the bounding box geometry, specified by the quadruple G = (x 0 , y 0 , x 1 , y 1 ) and A holds the recognized font attributes.Lines (also referred to as text lines) are an initially built aggregation of words that serves as auxiliary structure for all following procedures! They are described as quadruples L = (W 0 , succ, G, A), specifying a sorted list of words (with W 0 naming the first word and succ the appropriate successor function), the bounding box G and the attributes A = (linenumber, rownumber, spc len) that hold the unique linenumber, the logical rownumber (see Sect. 3.5) and average space width.Blocks are dynamic aggregations of words. Note that blocks are not aggregations of lines! The initial word-to-block mapping is made by the central clustering algorithm but is changed by the various postprocessing steps. Blocks thus keep the main segmentation information.Like lines, blocks are described as quadruple B = (W 0 , succ, G, A). The additional block attributes A = (type, justif ication, height, nmbwords) describe the classification into type 1 and 2 (see Sect. 3.2), the justification, height (as number of lines) and number of words in the block. DocumentThe document is defined as quadruple D = (B 0 , succ, G, A), where B 0 and succ specify a sorted list of blocks. The successor function succ doc (s) is defined for words, lines and blocks. The attributes A = (l spc) contain the average linespacing.Significantly large distances between adjacent lines cause the construction of dummy lines between them. This prevents the succ doc (line) function from bridging large line spacings as between paragraphs.In our notation, objects of higher level instances are denoted in a functional way: block(w x ) stands for the associated block of...

Image-based logical document structure recognition

“…OCR and rule based text analysis are helpful in recognition of specific text blocks, such as: captions, abstracts, document authors, etc. Tables without grid lines (plain text) are recognized by the T-Recs algorithm [18]. …”

“…Recognition of tables without grid lines These table structures are recognized using the T-Recs table recognition and analysis system [18]. T-Recs realises the bottomup clustering of word segments and does not apply any other top-down specific techniques (separator detection).…”

Disentangling the Structure of Tables in Scientific Literature

“…For example, table tags exist in HTML, but they are often used for formatting web page layout. Previous work focused on detecting tables from PDF, HTML and ASCII documents using Optical Character Recognition [13], machine learning algorithms such as C4.5 decision trees [17] or SVM [22,19], and heuristics [26].…”

A table-form extraction with artefact removal

“…Several studies have been presented on table recognition [1], [3], [6], [7], [8], [15], [16]. To reduce the complexity of the problem, some of them use tables composed by perfect horizontal and vertical line segments.…”