Catálogo de publicaciones - libros

This paper describes a Neural Network (NN) approach for logical document structure extraction. In this NN architecture, called Transparent Neural Network (TNN), the document structure is stretched along the layers, allowing an interpretation decomposition from physical (NN input) to logical (NN output) level. The intermediate layers represent successive interpretation steps. Each neuron is apparent and associated to a logical element. The recognition proceeds by repetitive perceptive cycles propagating the information through the layers. In case of low recognition rate, an enhancement is achieved by error backpropagation leading to correct or pick up a more adapted input feature subset. Several feature subsets are created using a modified filter method. The first experiments performed on scientific documents are encouraging.

We present in this paper a system for converting PDF legacy documents into structured XML format. This conversion system first extracts the different streams contained in PDF files (text, bitmap and vectorial images) and then applies different components in order to express in XML the logically structured documents. Some of these components are traditional in Document Analysis, other more specific to PDF. We also present a graphical user interface in order to check, correct and validate the analysis of the components. We eventually report on two real user cases where this system was applied on.

Accessing the structured content of PDF document is a difficult task, requiring pre-processing and reverse engineering techniques. In this paper, we first present different methods to accomplish this task, which are based either on document image analysis, or on electronic content extraction. Then, XCDF, a canonical format with well-defined properties is proposed as a suitable solution for representing structured electronic documents and as an entry point for further researches and works. The system and methods used for reverse engineering PDF document into this canonical format are also presented. We finally present current applications of this work into various domains, spacing from data mining to multimedia navigation, and consistently benefiting from our canonical format in order to access PDF document content and structures.

In this paper, a reliable and efficient structural analysis method for mathematical formulae is proposed for practical mathematical OCR. The proposed method consists of three steps. In the first step, a fast structural analysis algorithm is performed on each mathematical formula to obtain a tree representation of the formula. This step generally provides a correct tree representation but sometimes provides an erroneous representation. Therefore, the tree representation is verified by the following two steps. In the second step, the result of the analysis step, (i.e., a tree representation) is converted into a one-dimensional representation. The third step is a verification step where the one-dimensional representation is parsed by a formula description grammar, which is a context-free grammar specialized for mathematical formulae. If the one-dimensional representation is not accepted by the grammar, the result of the analysis step is detected as an erroneous result and alarmed to OCR users. This three-step organization achieves reliable and efficient structural analysis without any two-dimensional grammars.

The shift of interest to web tables in HTML and PDF files, coupled with the incorporation of table analysis and conversion routines in commercial desktop document processing software, are likely to turn table recognition into more of a systems than an algorithmic issue. We illustrate the transition by some actual examples of web table conversion. We then suggest that the appropriate target format for table analysis, whether performed by conventional customized programs or by off-the-shelf software, is a representation based on the abstract table introduced by X. Wang in 1996. We show that the Wang model is adequate for some useful tasks that prove elusive for less explicit representations, and outline our plans to develop a semi-automated table processing system to demonstrate this approach. Screen-snaphots of a prototype tool to allow table mark-up in the style of Wang are also presented.

An artefact identification method for handwritten filled table-forms is presented. Artefacts in table-forms are smudges and overlaps between handwritten data and line segments which increase the complexity of table-form interpretation. After reviewing some knowledge-based methods, a novel artefact identification method to improve table-form interpretation is presented. The proposed method aims to detect, identify and remove table-form artefacts with little use of previous knowledge. Experiments show the significance of using the proposed artefact identification method to improve table-form interpretation rates.

In this paper we present a text independent on-line writer identification system based on Gaussian Mixture Models (GMMs). This system has been developed in the context of research on Smart Meeting Rooms. The GMMs in our system are trained using two sets of features extracted from a text line. The first feature set is similar to feature sets used in signature verification systems before. It consists of information gathered for each recorded point of the handwriting, while the second feature set contains features extracted from each stroke. While both feature sets perform very favorably, the stroke-based feature set outperforms the point-based feature set in our experiments. We achieve a writer identification rate of 100% for writer sets with up to 100 writers. Increasing the number of writers to 200, the identification rate decreases to 94.75%.

This paper presents a study of 25 structural features extracted from samples of grapheme ‘th’ that correspond to features commonly used by forensic document examiners. Most of the features are extracted using vector skeletons produced by a specially developed skeletonisation algorithm. The methods of feature extraction are presented along with the results. Analysis of the usefulness of the features was conducted and three categories of features were identified: indispensable, partially relevant and irrelevant for determining the authorship of genuine unconstrained handwriting. The division was performed based on searching the optimal feature sets using the wrapper method. A constructive neural network was used as a classifier and a genetic algorithm was used to search for optimal feature sets. It is shown that structural micro features similar to those used in forensic document analysis do possess discriminative power. The results are also compared to those obtained in our preceding study, and it is shown that use of the vector skeletonisation allows both extraction of more structural features and improvement the feature extraction accuracy from 87% to 94%.

This paper describes a method of producing segmentation point candidates for on-line handwritten Japanese text by a support vector machine (SVM) to improve text recognition. This method extracts multi-dimensional features from on-line strokes of handwritten text and applies the SVM to the extracted features to produces segmentation point candidates. We incorporate the method into the segmentation by recognition scheme based on a stochastic model which evaluates the likelihood composed of character pattern structure, character segmentation, character recognition and context to finally determine segmentation points and recognize handwritten Japanese text. This paper also shows the details of generating segmentation point candidates in order to achieve high discrimination rate by finding the combination of the segmentation threshold and the concatenation threshold. We compare the method for segmentation by the SVM with that by a neural network using the database HANDS-Kondate_t_bf-2001-11 and show the result that the method by the SVM bring about a better segmentation rate and character recognition rate.

In this paper, we describe a bi-gram driven method for automatic reading of Chinese handwritten mails. In destination address block (DAB) location, text lines are first extracted by connected components analysis. Each candidate line is segmented and recognized by our holistic method, which incorporates mail layout features, recognition confidence and context cost. All these are also taken into consideration to identify the DABs from the candidate text lines. Based on them, street address line and organization name line are determined. At last step, edit distance based string matching is performed against given databases. We also discuss the pretreatment to deal with Chinese address databases consisted of a large amount of vocabularies in order to generate keywords for fast indexing during matching. Detailed experiment results on handwritten mail samples are given in the last section.