Search

International audience; This paper proposes an entity local structure comparison approach based on inexact subgraph matching. The comparison results are used for mislabeling correction in the local structure. The latter represents a set of entity attribute labels which are physically close in a document image. It is modeled by an attributed graph describing the content and presentation features of the labels by the nodes and the geometrical features by the arcs. A local structure graph is matched with a structure model which represents a set of local structure model graphs. The structure model is initially built using a set of well chosen local structures based on a graph clustering algorithm and is then incrementally updated. The subgraph matching adopts a specific cost function that integrates the feature dissimilarities. The matched model graph is used to extract the missed labels, prune the extraneous ones and correct the erroneous label fields in the local structure. The evaluation of the structure comparison approach on 525 local structures extracted from 200 business documents achieves about 90% for recall and 95% for precision. The mislabeling correction rates in these local structures vary between 73% and 100%.

This paper presents an approach for spotting textual fields in colored forms. We proceed by locating these fields thanks to their neighboring context which is modeled with a structural representation. First, informative zones are extracted. Second, forms are represented by graphs in which nodes represent colored rectangles while edges represent neighboring links. Finally, the context of the queried region of interest is modeled as a graph. Subgraph isomorphism is applied in order to locate this ROI in the structural representation of a whole document. Evaluated on a 130-document image dataset, experimental results show up that our approach is efficient and that the requested information is found even if its position is changed., Cet article présente une approche permettant la localisation de champs de saisie sur des images couleur de formulaires. Ces champs sont localisés grâce à une modélisation structurelle représentant leur contexte. Dans un premier temps, les zones informatives sont ex-traites. Les formulaires sont ensuite représentés par des graphes au sein desquels les noeuds représentent des rectangles de couleur uniforme tandis que les arcs modélisent les relations de voisinage. Finalement, le voisinage de la région d'intérêt à localiser est également représenté par un graphe. Une recherche d'isomorphisme de sous graphe vise à localiser le graphe modélisant le voisinage de la région d'intérêt au sein de la représentation structurelle du document cible. Une expérimentation est réalisée sur une base de 130 images de document. Les résultats montrent l'efficacité de la méthode même si la position de la région d'intérêt est variable.

International audience; —Administrative document images are usually processed in black and white what generates many problems due to the errors related to the binarization. Besides all semantic information provided by the color is lost. Document images have a rich and highly variable content. The presence of false colors and artefacts introduced by the scanning and the compression alter the segmentation of the regions. Problems arise when there is no correspondence between the point clouds which are detected in a color space and the real regions of an image. In order to help the segmentation, we propose the extraction of the main colors of an image as a set of binary layers. Due to the industrial context, our approach has to run unsupervised on a generic dataset of color administrative documents. The originality of this approach is the use of a multiresolution analysis to detect the number of colors automatically. At a low resolution, a set of local regions is obtained thanks to a SLIC-based approach which takes into account the structure of documents and which combines both colorimetric information and spatial information. Then, a merging stage is applied on each resolution separately based on the colors which have been extracted at a lower resolution. This contribution can both feed the traditional process and exploit colorimetric information.

International audience; —This paper presents an approach for spotting tex-tual fields in commercial and administrative colored forms. We proceed by locating these fields thanks to their neighboring context which is modeled with a structural representation. First, informative zones are extracted. Second, forms are represented by graphs. In these graphs, nodes represent colored rectangular shapes while edges represent neighboring relations. Finally, the neighboring context of the queried region of interest is modeled as a graph. Subgraph isomorphism is applied in order to locate this ROI in the structural representation of a whole document. Evaluated on a 130-document image dataset, experimental results show up that our approach is efficient and that the requested information is found even if its position is changed.

Given the number of existing classification algorithms, finding the most appropriate for classifying a new corpus is a difficult task. Meta-classification appears today very useful to help to determine, by using past experiences, what should be the most suitable algorithm compared to our corpus. The underlying idea is that "if an algorithm was particularly suitable for a corpus, it should have the same behavior on a quite similar corpus.". In this paper, we propose new meta-descriptors based on the concept of similarity to improve the meta-classification step. Conducted experiments on real dataset show the relevance of our new meta-descriptors., Devant le nombre d'algorithmes de classification existants, trouver l'algorithme qui sera le plus adapté pour classer un corpus de documents est une tâche difficile. La méta-classification apparaît aujourd'hui très utile pour aider à déterminer, en fonction des expé-riences passées, quel devrait être l'algorithme le plus pertinent par rapport à notre corpus. L'idée sous jacente est que "si un algorithme s'est montré particulièrement adapté pour un corpus , il devrait avoir le même comportement sur un corpus assez similaire". Dans cet article, nous proposons de nouveaux méta-descripteurs reposant sur les notions de similarités pour améliorer l'étape de méta-classification. Les expérimentations menées sur différents jeux de données réelles montrent la pertinence de nos nouveaux descripteurs.

Every day, classification is omnipresent and unconscious. For example in the process of decision when faced with something (an object, an event, a person), we will instinctively think of similar elements in order to adapt our choices and behaviors. This storage in a particular category is based on past experiences and characteristics of the element. The largest and the most accurate will be experiments, the most relevant will be the decision. It is the same when we need to categorize a document based on its content. For example detect if there is a children’s story or a philosophical treatise. This treatment is of course more effective if we have a large number of works of these two categories and if books had a large number of words.In this thesis we address the problem of decision making precisely when we have few learning documents and when the documents had a limited number of words. For this we propose a new approach based on new weights. It enables us to accurately determine the weight to be given to the words which compose the document. To optimize treatment, we propose a configurable approach. Five parameters make our adaptable approach, regardless of the classification given problem. Numerous experiments have been conducted on various types of documents in different languages and in different configurations. According to the corpus, they highlight that our proposal allows us to achieve superior results in comparison with the best approaches in the literature to address the problems of small dataset.The use of parameters adds complexity since it is then necessary to determine optimital values. Detect the best settings and best algorithms is a complicated task whose difficulty is theorized through the theorem of No-Free-Lunch. We treat this second problem by proposing a new meta-classification approach based on the concepts of distance and semantic similarities. Specifically we propose new meta-features to deal in the context of classification of documents. This original approach allows us to achieve similar results with the best approaches to literature while providing additional features.In conclusion, the work presented in this manuscript has been integrated into various technical implementations, one in the Weka software, one in a industrial prototype and a third in the product of the company that funded this work.; Au quotidien, le réflexe de classifier est omniprésent et inconscient. Par exemple, dans le processus de prise de décision où face à un élément (un objet, un événement, une personne) nous allons instinctivement chercher à rapprocher cet élément d’autres similaires afin d’adapter nos choix et nos comportements. Cette association à telle ou telle catégorie repose sur les expériences passées et les caractéristiques de l’élément. Plus les expériences seront nombreuses et les caractéristiques détaillées, plus fine et pertinente sera la décision. Il en est de même lors- qu’il nous faut catégoriser un document en fonction de son contenu. Par exemple détecter s’il s’agit d’un conte pour enfants ou d’un traité de philosophie. Ce traitement est bien sûr d’autant plus efficace si nous possédons un grand nombre d’ouvrages de ces deux catégories et que l’ouvrage à classifier possède un nombre important de mots.Dans cette thèse, nous nous intéressons à la problématique de la prise de décision lorsque nous disposons de peu de documents d’apprentissage et que le document possède un nombre de mots limité. Nous proposons pour cela une nouvelle approche qui repose sur de nouvelles pondérations. Elle nous permet de déterminer avec précision l’importance à accorder aux mots composant le document.Afin d’optimiser les traitements, nous proposons une approche paramétrable. Cinq paramètres rendent notre système adaptable, quel que soit le problème de classification donné. De très nombreuses expérimentations ont été menées sur différents types de documents qui sont de langues variées et en appliquant différentes configurations. Selon les corpus, elles mettent en évidence que notre proposition permet d’obtenir des résultats supérieurs en comparaison avec les meilleures approches de la littérature pour traiter de petits volumes de données.L’utilisation de paramètres introduit une complexité supplémentaire puisqu’il faut alors déterminer les valeurs optimales. Détecter les meilleurs paramètres et algorithmes est une tâche compliquée dont la difficulté est théorisée à travers le théorème du No-Free-Lunch. Nous traitons cette seconde problématique en proposant une nouvelle approche de méta-classification reposant sur les notions de distances et de similarités. Plus précisément, nous proposons de nouveaux méta-descripteurs adaptés dans un contexte de classification de documents. Cette approche originale nous permet d’obtenir des résultats similaires aux meilleures approches de la littérature tout en offrant des qualités supplémentaires.

International audience; More and more data come with contextual information describing the circumstances of their acquisition. While the frequent pattern mining literature offers a lot of approaches to handle and extract interesting patterns in data, little effort has been dedicated to relevantly handling such contextual information during the mining process. In this paper we propose a generic formulation of the contextual frequent pattern mining problem and provide the CFPM algorithm to mine frequent patterns that are representative of a context. This approach is generic w.r.t. the pattern language (e.g., itemsets, sequential patterns, subgraphs, etc.) and therefore is applicable in a wide variety of use cases. The CFPM method is experimented on real datasets with three different pattern languages to assess its performances and genericity.

1st International Workshop on Interactions between Data Mining and Natural Language Processing; International audience; Regular expressions (REs) are a widely used tool when con- sidering textual data. They are, for instance, commonly used as search templates, or to check the validity of a piece of text w.r.t. some formatting requirements. Such applications typically rely on very small amounts of hand-crafted REs. However, automatically constructing REs matching with a large amount of string examples (e.g., a set of attribute values in a database) would help characterizing the formatting rules underlying these strings with no a priori knowledge, and would allow envisaging new RE-based applications. We propose to formulate the problem of dis- covering such REs as a frequent pattern mining problem.

International audience; —This paper addresses the problem of machine printed and handwritten text separation in real noisy documents. We have proposed in a previous work a robust separation system relying on a proximity string segmentation algorithm. The extracted pseudo-lines and pseudo-words are used as basic blocks for classification. A multi-class support vector machine (SVM) with Gaussian kernel associates first an appropriate label to each pseudo-word. Then, the local neighborhood of each pseudo-word is studied in order to propagate the context and correct the classification errors. In this work, we first propose to model the separation problem by conditional random fields considering the horizontal neighborhood. As the considered neighborhood is too local to solve certain error cases, we have enhanced this method by using a more global context based on class dominance in the pseudo-line. The method has been evaluated on business documents. It separates handwritten and printed text with better scores (99.1% and 99.2% respectively), contrary to noise which is very random in these documents (90.1%).

International audience; In text classification, providing an efficient classifier even if the number of documents involved in the learning step is small remains an important issue. In this paper we evaluate the performance of traditional classification methods to better evaluate their limitation in the learning phase when dealing with small amount of documents. We thus propose a new way for weighting features which are used for classifying. These features have been integrated in two well known classifiers: Class-Feature-Centroid and Naïve Bayes, and evaluations have been performed on two real datasets. We have also investigated the influence on parameters such as number of classes, documents or words in the classification. Experiments have shown the efficiency of our proposal relatively to state of the art classification methods. Either with a very few amount of data or with a small number of features that can be extracted from poor content documents, we show that our approach performs well.

International audience; Un des défis actuels dans le domaine de la classification supervisée de documents est de pouvoir produire un modèle fiable à partir d'un faible volume de données. Avec un volume conséquent de données, les classifieurs fournissent des résultats satisfaisants mais les performances sont dégradées lorsque celui-ci diminue. Nous proposons, dans cet article, de nouvelles méthodes de pondérations résistant à une diminution du volume de données. Leur efficacité, évaluée en utilisant des algorithmes de classification supervisés existants (Naive Bayes et Class-Feature-Centroid) sur deux corpus différents, est supérieure à celle des autres algorithmes lorsque le nombre de descripteurs diminue. Nous avons étudié en parallèle les paramètres influençant les différentes approches telles que le nombre de classes, de documents ou de descripteurs.

International audience; RÉSUMÉ. Cet article propose une approche de segmentation supervisée de flux de documents. L'approche traite le flux de documents comme une suite de paires de pages et étudie la relation qui existe entre elles pour déceler une continuité de documents ou une rupture. Dans un premier temps, des descripteurs sont extraits des pages et une approche est proposée pour fusionner ces descripteurs en un seul vecteur qui modélise la relation entre les paires de pages. Cette représentation est fournie à un classifieur binaire qui la classifie comme étant une rupture (synonyme de segmentation) ou une continuité. Dans le cas d'une rupture, nous considérons que nous avons atteint la limite d'un document complet et l'analyse du flux continue en commençant par un nouveau document. En cas d'une continuité, les deux pages sont considérées comme appartenant à un même document. S'il y a une incertitude sur la classe de la limite, un rejet est décidé et les pages analysées jusqu'à ce point sont considérées comme un « fragment » on réalise ici une sur-segmentation. Cette classification donne de bons résultats approchant 90% sur certains documents, ce qui est élevé à ce niveau du système. ABSTRACT. This paper proposes a document flow supervised segmentation approach. Our algorithm treats the flow of documents as couples of consecutive pages and examines the relationship that exists between them in order to present a document continuity or rupture. In a first step, descriptors are extracted from the pages and an approach is proposed to merge these descriptors into a single vector that models the relationship between pairs of pages. This representation is provided to a binary classifier that classifies it as either a rupture (synonymous with segmentation) or continuity. In case of a rupture, we consider that the limit of a complete document has been reached and the stream analysis continues by starting a new document. In case of continuity, the two pages are considered to belong to the same document. If there is an uncertainty on the class of the limit, a rejection is decided and the pages analyzed until this point are considered as a "fragment" and an over-segmentation is applied. The classification provides good results approaching 90% on certain documents, which is high at this level of the system. MOTS-CLÉS : Segmentation de flux de documents, descripteurs textuels, classification.

International audience; This chapter reports our knowledge about the analysis and recognition of scanned administrative documents. Regarding essentially the administrative paper flow with new and continuous arrivals, all the conventional techniques reserved to static databases modeling and recognition are doomed to failure. For this purpose, a new technique based on the experience was investigated giving very promising results. This technique is related to the case-based reasoning already used in data mining and various problems of machine learning. After the presentation of the context related to the administrative document flow and its requirements in a real time processing, we present a case based reasonning for invoice processing. The case corresponds to the co-existence of a problem and its solution. The problem in an invoice corresponds to a local structure such as the keywords of an address or the line patterns in the amounts table, while the solution is related to their content. This problem is then compared to a document case base using graph probing. For this purpose, we proposed an improvement of an already existing neural network called Incremental Growing Neural Gas

International audience; This paper summarizes some techniques proposed for off-line Arabic word recognition. The point of view developed here concerns the human reading favoring an interactive mechanism between global memorization and local checking making easier the recognition of complex scripts as Arabic. According to this consideration, some specific papers are analyzed and their strategies commented

International audience; In this paper, an approach for adaptive region segmentation of mail-order forms for high volume application is described. Regions are first identified through a selection of their anchor points described by a constraint graph, illustrating their typographic aspects in the nodes, and their topographical relationships in the arcs. Then the identification of the actual anchor points is performed from a list of textual candidates, using the Arc Consistency Algorithm (AC4). Finally, some contextual heuristics are investigated for properly delimiting the regions. The originality of this approach lies mainly in the absence of a rigid a priori model, replaced by a simply and reliable association of anchor points. The constraint graph used for their description can be easily derived from a general logical definition of their content. Experimental results are overall encouraging and the methodology integration is under execution for commercialization.

International audience; The present paper is focused on information extraction from key fields of invoices using two different methods based on sequence labeling. Invoices are semi-structured documents in which data can be located based on the context. Common information extraction systems are model-driven, using heuristics and lists of trigger words curated by domain experts. Their performances are generally high on documents they have been trained for but processing new templates often requires new manual annotations, which is tedious and time-consuming to produce. Recent works on deep learning applied to business documents claimed a gain in terms of time and performance. While these systems do not need manual curation, they nevertheless require a large amount of data to achieve good results. In this paper, we present a series of experiments using neural networks approaches to study the trade-off between data requirements and performance in the extraction of information from key fields of invoices (such as dates, document numbers, types, amounts...). The main contribution of this paper is a system that achieves competitive results using a small amount of data compared to the state-of-the-art systems that need to be trained on large datasets, that are costly and impractical to produce in real-world applications.

One major drawback of state of the art Neural Networks (NN)-based approaches for document classification purposes is the large number of training samples required to obtain an efficient classification. The minimum required number is around one thousand annotated documents for each class. In many cases it is very difficult, if not impossible, to gather this number of samples in real industrial processes. In this paper, we analyse the efficiency of NN-based document classification systems in a sub-optimal training case, based on the situation of a company’s document stream. We evaluated three different approaches, one based on image content and two on textual content. The evaluation was divided into four parts: a reference case, to assess the performance of the system in the lab; two cases that each simulate a specific difficulty linked to document stream processing; and a realistic case that combined all of these difficulties. The realistic case highlighted the fact that there is a significant drop in the efficiency of NN-Based document classification systems. Although they remain efficient for well represented classes (with an over-fitting of the system for those classes), it is impossible for them to handle appropriately less well represented classes. NN-Based document classification systems need to be adapted to resolve these two problems before they can be considered for use in a company’s document stream.

International audience; This paper describes the ICPR2018 fraud detection contest, its data set, evaluation methodology, as well as the different methods submitted by the participants to tackle the predefined tasks. Forensics research is quite a sensitive topic. Data are either private or unlabeled and most of related works are evaluated on private datasets with a restricted access. This restriction has two major consequences: results cannot be reproduced and no benchmarking can be done between every approach. This contest was conceived in order to address these drawbacks. Two tasks were proposed: detecting documents containing at least one forgery in a flow of documents and spotting and localizing these forgeries within documents. An original dataset composed of images and texts of French receipts was provided to participants. The results they obtained are presented and discussed.

International audience; The aim of this paper is to introduce a new dataset initially created to work on fraud detection in documents. This dataset is composed of 1969 images of receipts and the associated OCR result for each. The article details the dataset and its interest for the document analysis community. We indeed share this dataset with the community as a benchmark for the evaluation of fraud detection approaches.

International audience; In this paper, we present a model-based document information content extraction approach and perform in-depth evaluation based on clients' relevance. Real-world users i.e., clients first provide a set of key fields from the document image which they think are important. These are used to represent a graph where nodes (i.e., fields) are labelled with dynamic semantics including other features and edges are attributed with spatial relations. Such an attributed relational graph (ARG) is then used to mine similar graphs from a document image that are used to reinforce or update the initial graph iteratively each time we extract them, in order to produce a model. Models therefore, can be employed in the absence of clients. We have validated the concept and evaluated its scientific impact on real-world industrial problem, where table extraction is found to be the best suited application.

International audience; In this paper, we address a client-driven approach to automatically extract information content within the table in document images. We start with a graph-based representation of a set of key-fields selected by clients and perform graph mining in a document in order to learn them to produce a model. Such models are aimed to use to extract information content in the absence of clients. To avoid NP-hard general problem, our graph matching is based on relation assignment to see whether pairs of nodes are semantically identical. We have validated the concept by using a real-world industrial problem.

International audience; The goal of the project is to extract content within table in document images based on learnt patterns. Real-world users i.e., clients first provide a set of key fields within the table which they think are important. These are first used to represent the graph where nodes are labelled with semantics including other features and edges are attributed with relations. Attributed relational graph (ARG) is then employed to mine similar graphs from a document image. Each mined graph will represent an item within the table, and hence a set of such graphs will compose a table. We have validated the concept by using a real-world industrial problem.