The T-recs Approach for Table Structure Recognition and Table Border Determination

We present a snapshot of the ongoing research in the eld of table structure recognition and analysis. The prototypical T-Recs system (Table RECognition System) relies on the word level layout (bounding box geometry) as primary input. It moreover considers the textual information and potentially available delineations as further input. This article resumes the basic ideas and system features as described in [1] and [2] (downloadable from our demo page: www.dfki.uni-kl.de/ kieni/t recs/). This page also allows to interactivly load and change some prede ned documents which demonstrate the main features and strengths of the approach. Next, we will sketch some problems which encounter when applying T-Recs to business letters like o ers, invoices etc. and discuss appropriate solutions for that problem. 1 SUMMARY OF CURRENT SYSTEM Based on the word level layout information which can be derived from either OCR documents (e.g. the Xerox XDOC-format) or plain ASCII les by using a built-in preprocessor, the TRecs system performs a direct bottom-up clustering of word segments to blocks (not the conventional word-line-block order!) and thus transforms the document representation from the word level layout to a partial logical representation (see Figure 1). Table Structures Blocksegments Wordsegments Structural Analysis Segmentation Logic