Search

This paper presents a deep learning approach for image retrieval and pattern spotting in digital collections of historical documents. First, a region proposal algorithm detects object candidates in the document page images. Next, deep learning models are used for feature extraction, considering two distinct variants, which provide either real-valued or binary code representations. Finally, candidate images are ranked by computing the feature similarity with a given input query. A robust experimental protocol evaluates the proposed approach considering each representation scheme (real-valued and binary code) on the DocExplore image database. The experimental results show that the proposed deep models compare favorably to the state-of-the-art image retrieval approaches for images of historical documents, outperforming other deep models by 2.56 percentage points using the same techniques for pattern spotting. Besides, the proposed approach also reduces the search time by up to 200x and the storage cost up to 6,000x when compared to related works based on real-valued representations., 7 pages

Pattern spotting consists of locating different instances of a given object (i.e. an image query) in a collection of historical document images. These patterns may vary in shape, size, color, context and even style because they are hand-drawn, which makes pattern spotting a difficult task. To tackle this problem, we propose a Convolutional Neural Network (CNN) approach based on Feature Pyramid Networks (FPN) as the feature extractor of our system. Using FPN allows to extract descriptors of local regions of the documents to be indexed and queries, at multiple scales with just a single forward pass. Experiments conducted on DocExplore dataset show that the proposed system improves mAP by 73% (from 0.157 to 0.272) in pattern localization compared with state-of-the-art results, even when the feature extractor is not trained with domain-specific data. Memory requirement and computation time are also decreased since the descriptor dimension used for distance computation is reduced by a factor of 16.

Multi-view learning is a learning task in which data is described by several concurrent representations. Its main challenge is most often to exploit the complementarities between these representations to help solve a classification/regression task. This is a challenge that can be met nowadays if there is a large amount of data available for learning. However, this is not necessarily true for all real-world problems, where data are sometimes scarce (e.g. problems related to the medical environment). In these situations, an effective strategy is to use intermediate representations based on the dissimilarities between instances. This work presents new ways of constructing these dissimilarity representations, learning them from data with Random Forest classifiers. More precisely, two methods are proposed, which modify the Random Forest proximity measure, to adapt it to the context of High Dimension Low Sample Size (HDLSS) multi-view classification problems. The second method, based on an Instance Hardness measurement, is significantly more accurate than other state-of-the-art measurements including the original RF Proximity measurement and the Large Margin Nearest Neighbor (LMNN) metric learning measurement., accepted to ICPR 2020 (22/06/2020)

Pattern spotting consists of searching in a collection of historical document images for occurrences of a graphical object using an image query. Contrary to object detection, no prior information nor predefined class is given about the query so training a model of the object is not feasible. In this paper, a convolutional neural network approach is proposed to tackle this problem. We use RetinaNet as a feature extractor to obtain multiscale embeddings of the regions of the documents and also for the queries. Experiments conducted on the DocExplore dataset show that our proposal is better at locating patterns and requires less storage for indexing images than the state-of-the-art system, but fails at retrieving some pages containing multiple instances of the query., 6 pages, 9 figures

This paper presents a novel approach for image retrieval and pattern spotting in document image collections. The manual feature engineering is avoided by learning a similarity-based representation using a Siamese Neural Network trained on a previously prepared subset of image pairs from the ImageNet dataset. The learned representation is used to provide the similarity-based feature maps used to find relevant image candidates in the data collection given an image query. A robust experimental protocol based on the public Tobacco800 document image collection shows that the proposed method compares favorably against state-of-the-art document image retrieval methods, reaching 0.94 and 0.83 of mean average precision (mAP) for retrieval and pattern spotting (IoU=0.7), respectively. Besides, we have evaluated the proposed method considering feature maps of different sizes, showing the impact of reducing the number of features in the retrieval performance and time-consuming., Comment: Accepted for IJCNN 2019

Cancer diagnosis and treatment often require a personalized analysis for each patient nowadays, due to the heterogeneity among the different types of tumor and among patients. Radiomics is a recent medical imaging field that has shown during the past few years to be promising for achieving this personalization. However, a recent study shows that most of the state-of-the-art works in Radiomics fail to identify this problem as a multi-view learning task and that multi-view learning techniques are generally more efficient. In this work, we propose to further investigate the potential of one family of multi-view learning methods based on Multiple Classifiers Systems where one classifier is learnt on each view and all classifiers are combined afterwards. In particular, we propose a random forest based dynamic weighted voting scheme, which personalizes the combination of views for each new patient for classification tasks. The proposed method is validated on several real-world Radiomics problems., 10 pages

Breast cancer is one of the most common types of cancer and leading cancer-related death causes for women. In the context of ICIAR 2018 Grand Challenge on Breast Cancer Histology Images, we compare one handcrafted feature extractor and five transfer learning feature extractors based on deep learning. We find out that the deep learning networks pretrained on ImageNet have better performance than the popular handcrafted features used for breast cancer histology images. The best feature extractor achieves an average accuracy of 79.30%. To improve the classification performance, a random forest dissimilarity based integration method is used to combine different feature groups together. When the five deep learning feature groups are combined, the average accuracy is improved to 82.90% (best accuracy 85.00%). When handcrafted features are combined with the five deep learning feature groups, the average accuracy is improved to 87.10% (best accuracy 93.00%).

Today, medical image analysis papers require solid experiments to prove the usefulness of proposed methods. However, experiments are often performed on data selected by the researchers, which may come from different institutions, scanners, and populations. Different evaluation measures may be used, making it difficult to compare the methods. In this paper, we introduce a dataset of 7909 breast cancer histopathology images acquired on 82 patients, which is now publicly available from http://web.inf.ufpr.br/vri/breast-cancer-database . The dataset includes both benign and malignant images. The task associated with this dataset is the automated classification of these images in two classes, which would be a valuable computer-aided diagnosis tool for the clinician. In order to assess the difficulty of this task, we show some preliminary results obtained with state-of-the-art image classification systems. The accuracy ranges from 80% to 85%, showing room for improvement is left. By providing this dataset and a standardized evaluation protocol to the scientific community, we hope to gather researchers in both the medical and the machine learning field to advance toward this clinical application.

International audience; In this paper, we introduce a new Random Forest (RF) induction algorithm called Dynamic Random Forest (DRF) which is based on an adaptative tree induction procedure. The main idea is to guide the tree induction so that each tree will complement as much as possible the existing trees in the ensemble. This is done here through a resampling of the training data, inspired by boosting algorithms, and combined with other randomization processes used in traditional RF methods. The DRF algorithm shows a significant improvement in terms of accuracy compared to the standard static RF induction algorithm.

In this paper, we propose a multi-objective optimization framework for SVM hyperparameters tuning. The key idea is to manage a population of classifiers optimizing both False Positive and True Positive rates rather than a single classifier optimizing a scalar criterion. Hence, each classifier in the population optimizes a particular trade-off between the objectives. Within the context of two-class classification problems, our work introduces "the receiver operating characteristics (ROC) front concept" depicting a population of SVM classifiers as an alternative to the receiver operating characteristics (ROC) curve representation. The proposed framework leads to a noncost sensitive SVM training relying on the pool of classifiers. The comparison with a traditional scalar optimization technique based on an AUC criterion shows promising results on UCI datasets.

Writing and drawing tests are widely used in the clinical environment for the diagnosis of a variety of neuropsychological conditions. Conventional assessment of these tests involves the inspection by trained assessors of the completed patient response. This article describes the development of a computer-based framework for data capture, automated feature analysis, and result reporting for a range of drawing- and writing-based test batteries. In developing this framework, we have exploited the commonality between tasks while allowing for both flexibility in configuration across condition-specific testing requirements and extensibility for future test development. Using the two example clinical conditions of visuospatial neglect and dyspraxia, we illustrate the advantages of utilizing a computer-based analysis system, describe a structured approach to system implementation, and demonstrate the generality of this implementation for different conditions of interest, which extends to feature selection and design.

International audience; In this paper, we propose a query by string word spotting system able to extract arbitrary key-words in handwritten documents, taking both segmen-tation and recognition decisions at the line level. The system relies on the combination of a HMM line model made of keyword and non-keyword (filler) models, with a deep neural network (DNN) that estimates the state-dependent observation probabilities. Experiments are carried out on RIMES database, an unconstrained hand-written document database that is used for benchmark-ing different handwriting recognition tasks. The ob-tained results show the superiority of the proposed frame-work over the classical GMM-HMM and standard HMM hybrid architectures.

International audience

This paper deals with effective separation of foreground and background in low quality document images suffering from various types of degradations including scanning noise, aging effects, uneven background, or foreground, etc. The proposed algorithm shows an excellent adaptability to tackle with these problems of uneven illumination and local changes or nonuniformity in background and foreground colors. The approach is primarily designed for (not restricted to) processing of color documents but it works well in the gray scale domain too. Test document set considers samples (in color as well as in gray scale) of old historical documents including manuscripts of high importance. The data set used in this study consists of hundred images. These images are selected from different sources including image databases that had been scanned from working notebooks of famous writers who used to write with quill or pencil generating very low contrast between foreground and background. Evaluation of foreground extraction method has been judged by computing the accuracy of extracting handwritten lines and words from the test images. This evaluation shows that the proposed method can extract lines and words with accuracies of about 84% and 93%, respectively. Apart from this quantitative method, a qualitative evaluation is also presented to compare the proposed method with one popular technique for foreground/background separation in document images.

The aim of this paper is to assess the evolution in writing performance amongst typical pupils in primary education. More precisely, we propose ways of discovering groups of pupils sharing the same writing strategies during their primary education and methods for the temporal modeling of these pupils' writing strategies. For this purpose, online acquisition of writing and drawing tests have been performed three times during a period of one year for the same pupils under the same experimental conditions. A first approach, based on clustering, is applied to highlight clusters on a set of dynamic primitives chosen by an expert in the field of child development psychology. Results are presented by means of a comparative study between features of each group and writing tests. An analysis of within and between-strategies migration of pupils over time is also conducted to highlight pupils who change (or fail to change) their writing strategies during this period of one year. A second approach is used to model the problem by means of a probabilistic graphical model, i.e. a bayesian network. Expert knowledge partially determines the bayesian network structure, in which the writing strategy is represented by a hidden variable whose cardinality is estimated by the results of the clustering approach. By considering that each writing test is represented by its own (local) strategy and that there exists a global strategy which deals with each local strategy, we propose a Global Hierarchical Model. The results of our hierarchical model structured using real data highlight, among others, two global strategies that correspond to normo-writer pupils and more advanced normo-writers. A longitudinal and temporal study of the evolution of the pupils in these strategies shows that these two strategies are consistent.

This paper presents a complete numeral amount recognition module which is integrated in an automatic system aimed at reading all types of French checks. This module is combined with an automatic reading system of literal amounts. This complete working system, called LIREChèques, is developed by MATRA MS&I and is now in advanced test at SERINTEL, a pilot site. Two aspects of the numeral amount recognition system are particularly emphasized: the numeral recognition stage itself and the syntactic analysis stage. The numeral recognition module relies on a combination of two individual classifiers, the first one is based on concavity measurements, the second one on both statistical and structural features. The syntactic analysis, called syntactic/contextual analysis, is combined with contextual information to take into account the segmentation behaviour and the presence of literal entities in the numeral amount. We demonstrate that very good performances can be obtained on digits such as those extracted from numeral amounts since a substitution rate of 0.06% while still preserving a recognition rate of near 87% can be achieved. As for the syntactic/contextual analysis stage, results obtained on a test set (containing checks from more than 40 different banks and 15/ of typed checks, thus being a good representation of the real tests realized on site) show clearly that introduction of contextual information in association with syntactic analysis allows to process much more numeral amounts than a simple syntactic analysis and increases perceptibility of the recognition rate.

International audience; We propose a system designed to spot either words or patterns, based on a user made query. Employing a two stage approach, it takes advantage of the descriptive power of the Bag of Visual Words (BOVW) representation and the discriminative power of the proposed Longest Weighted Profile (LWP) algorithm. First, we try to identify the zones of images that share common characteristics with the query as summed up in a BOVW. Then, we filter these zones using the LWP introducing spatial constraints extracted from the query. We have validated our system on the George Washington handwritten document database for word spotting, and medieval manuscripts from the DocExplore project for pattern spotting.

International audience; One class classification is a binary classification task for which only one class of samples is available for learning. In some preliminary works, we have proposed One Class Random Forests (OCRF), a method based on a random forest algorithm and an original outlier generation procedure that makes use of classifier ensemble randomization principles. In this paper, we propose an extensive study of the behavior of OCRF, that includes experiments on various UCI public datasets and comparison to reference one class namely, Gaussian density models, Parzen estimators, Gaussian mixture models and One Class SVMs--with statistical significance. Our aim is to show that the randomization principles embedded in a random forest algorithm make the outlier generation process more efficient, and allow in particular to break the curse of dimensionality. One Class Random Forests are shown to perform well in comparison to other methods, and in particular to maintain stable performance in higher dimension, while the other algorithms may fail.

International audience; In this paper, we describe DocExplore, an integrated software suite centered on the handling of digitized documents with an emphasis on ancient manuscripts. This software suite allows the augmentation and exploration of ancient documents of cultural interest. Specialists can add textual and multimedia data and metadata to digitized documents through a graphical interface that does not require technical knowledge. They are helped in this endeavor by sophisticated document analysis tools that allows for instance to spot words or patterns in images of documents. The suite is intended to ease considerably the process of bringing locked away historical materials to the attention of the general public by covering all the steps from managing a digital collection to creating interactive presentations suited for cultural exhibitions. Its genesis and sustained development reside in a collaboration of archivists, historians and computer scientists, the latter being not only in charge of the development of the software, but also of creating and incorporating novel pattern recognition for document analysis techniques.

Recently, the in vivo imaging of pulmonary alveoli was made possible thanks to confocal microscopy. For these images, we wish to aid the clinician by developing a computer-aided diagnosis system, able to discriminate between healthy and pathological subjects. The lack of expertise currently available on these images has first led us to choose a generic approach, based on pixel-value description of randomly extracted subwindows and decision tree ensemble for classification (extra-trees). In order to deal with the great complexity of our images, we adapt this method by introducing a texture-based description of the subwindows, based on local binary patterns. We show through our experimental protocol that this adaptation is a promising way to classify fibered confocal fluorescence microscopy images. In addition, we introduce a rejection mechanism on the classifier output to prevent nondetection errors.

International audience; A novel imaging technique can now provide microscopic images of the distal lung in vivo, for which quantitative analysis tools need to be developed. In this paper, we present an image classification system that is able to discriminate between normal and pathological images. Different feature spaces for discrimination are investigated and evaluated using a Support Vector Machine. Best classification rates reach up to 90% and 95% on non-smoker and smoker groups, respectively. A feature selection process is also implemented, that allows us to gain some insight about these images. Whereas further tests on extended databases are needed, these first results indicate that efficient computer based automated classification of normal vs. pathological images of the distal lung is feasible.

International audience; This paper presents a complete system able to categorize handwritten documents, i.e. to classify documents according to their topic. The categorization approach is based on the detection of some discrim- inative keywords prior to the use of the well known tf-idf representation for document categorization. Two keyword extraction strategies are explored. The rst one proceeds to the recognition of the whole document. However, the performance of this strategy strongly de- creases when the lexicon size increases. The second strat- egy only extracts the discriminative keywords in the handwritten documents. This information extraction strategy relies on the integration of a rejection model (or anti-lexicon model) in the recognition system. Ex- periments have been carried out on an unconstrained handwritten document database coming from an indus- trial application concerning the processing of incoming mails. Results show that the discriminative keyword ex- traction system leads to better recall/precision trade- o s than the full recognition strategy. The keyword ex- traction strategy also outperforms the full recognition strategy for the categorization task.

International audience; We propose a new one-class classification method, called One Class Random Forest, that is able to learn from one class of samples only. This method, based on a random forest algorithm and an original outlier generation procedure, makes use of the ensemble learning mechanisms offered by random forest algorithms to reduce both the number of artificial outliers to generate and the size of the feature space in which they are generated. We show that One Class Random Forests perform well on various UCI public datasets in comparison to few other state-of-the-art one class classification methods (gaussian density models, Parzen estimators, gaussian mixture models and one-class SVMs).

International audience; In this paper, we introduce an alpha-numerical sequences extraction system (keywords, numerical fields or alpha-numerical sequences) in unconstrained handwritten documents. Contrary to most of the approaches presented in the literature, our system relies on a global handwriting line model describing two kinds of information : i) the relevant information and ii) the irrelevant information represented by a shallow parsing model. The shallow parsing of isolated text lines allows quick information extraction in any document while rejecting at the same time irrelevant information. Results on a public french incoming mails database show the efficiency of the approach.

International audience; Until recently, the alveolar region could not be investigated in-vivo. A novel technique, based on confocal microscopy, can now provide new images of the respiratory alveolar system, for which quantitative analysis tools must be developed, for diagnosis and follow up of pathological situations. In particular, we wish to aid the clinician by developing a computer-aided diagnosis system, able to discriminate between healthy and pathological subjects. This paper describes this system, in which images are first characterized through a 148-feature vector then classified by an SVM (Support Vector Machine). Experiments conducted on smoker and non smoker images show that the dimensionality of the feature vector can be reduced significantly without decreasing classification accuracy, and thus gaining some insight about the usefulness of features for medical diagnosis. These promising results allow us to consider interesting perspectives for this very challenging medical application.

International audience; In this paper, a new information extraction system by statistical shallow parsing in unconstrained handwritten documents is introduced. Unlike classical approaches found in the literature as keyword spotting or full document recognition, our approch relies on a strong and powerful global handwriting model. A entire text line is considered as an indivisible entity and is modeled with Hidden Markov Models. In this way, text line shallow parsing allows fast extraction of the relevant information in any document while rejecting at the same time irrelevant information. First results are promising and show the interest of the approach.

International audience; In this paper, we tackle the problem of model selection when misclassification costs are unknown and/or may evolve. Unlike traditional approaches based on a scalar optimization, we propose a generic multi-model selection framework based on a multi-objective approach. The idea is to automatically train a pool of classifiers instead of one single classifier, each classifier in the pool optimizing a particular trade-off between the objectives. Within the context of two-class classification problems, we introduce the "ROC front concept" as an alternative to the ROC curve representation. This strategy is applied to the multi-model selection of SVM classifiers using an evolutionary multi-objective optimization algorithm. The comparison with a traditional scalar optimization technique based on an AUC criterion shows promising results on UCI datasets as well as on a real-world classification problem.

International audience; This paper presents a new approach for the classification of pathological vs. healthy endomicroscopic images of the alveoli. These images, never seen before, require an adequate description. We investigate two types of feature vector for discrimination: a high-level feature vector based on visual analysis of the images, and a pixel-based, generic feature vector, based on Local Binary Patterns (LBP). Both feature sets are evaluated on state-of-the-art classifiers and an intensive study of the LBP parameters is conducted. Indeed best results are obtained with the LBP-based approach, with correct classification rates reaching up to 91.73\% and 97.85\% for non-smoking and smoking groups, respectively. Even though tests on extended databases are needed, first results are very encouraging for this difficult task of classifying endomicroscopic images of the distal lung.

In this paper we present our work on the Random Forest (RF) family of classification methods. Our goal is to go one step further in the understanding of RF mechanisms by studying the parametrization of the reference algorithm Forest-RI. In this algorithm, a randomization principle is used during the tree induction process, that randomly selects K features at each node, among which the best split is chosen. The strength of randomization in the tree induction is thus led by the hyperparameter K which plays an important role for building accurate RF classifiers. We have decided to focus our experimental study on this hyperparameter and on its influence on classification accuracy. For that purpose, we have evaluated the Forest-RI algorithm on several machine learning problems and with different settings of K in order to understand the way it acts on RF performance. We show that default values of K traditionally used in the literature are globally near-optimal, except for some cases for which they are all significatively sub-optimal. Thus additional experiments have been led on those datasets, that highlight the crucial role played by feature relevancy in finding the optimal setting of K .

In this paper we present a study on the Random Forest (RF) family of ensemble methods. From our point of view, a ”classical” RF induction process presents two main drawbacks : (i) the number of trees has to be a priori fixed (ii) trees are independently, thus arbitrarily, added to the ensemble due to the randomization principle. Hence, this kind of process offers no guarantee that all the trees will well cooperate into the same committee. In this work we thus propose to study the RF mechanisms that explain this cooperation by analysing, for particular subsets of trees called sub-forests, the link between accuracy and properties such as Strength and Correlation. We show that these properties, through the Correlation/Strengh ratio, should be taken into account to explain the sub-forest performance.

In this paper we present our work on the parametrization of Random Forests (RF), and more particularly on the number K of features randomly selected at each node during the tree induction process. It has been shown that this hyperparameter can play a significant role on performance. However, the choice of the value of K is usually made either by a greedy search that tests every possible value to choose the optimal one, either by choosing a priori one of the three arbitrary values commonly used in the literature. With this work we show that none of those three values is always better than the others. We thus propose an alternative to those arbitrary choices of K with a new ”push-button” RF induction method, called Forest-RK, for which K is not an hyperparameter anymore. Our experimentations show that this new method is at least as statistically accurate as the original RF method with a default K setting.

International audience; In this paper we present a study on the random forest (RF) family of ensemble methods. In a ldquoclassicalrdquo RF induction process a fixed number of randomized decision trees are inducted to form an ensemble. This kind of algorithm presents two main drawbacks : (i) the number of trees has to be fixed a priori (ii) the interpretability and analysis capacities offered by decision tree classifiers are lost due to the randomization principle. This kind of process in which trees are independently added to the ensemble, offers no guarantee that all those trees will cooperate effectively in the same committee. This statement rises two questions: are there any decision trees in a RF that provide the deterioration of ensemble performance? If so, is it possible to form a more accurate committee via removal of decision trees with poor performance? The answer to these questions is tackled as a classifier selection problem. We thus show that better subsets of decision trees can be obtained even using a sub-optimal classifier selection method. This proves that ldquoclassicalrdquo RF induction process, for which randomized trees are arbitrary added to the ensemble, is not the best approach to produce accurate RF classifiers. We also show the interest in designing RF by adding trees in a more dependent way than it is traditionally done in ldquoclassicalrdquo RF induction algorithms.

International audience; In the Pattern Recognition field, growing interest has been shown in recent years for Multiple Classifier Systems and particularly for Bagging, Boosting and Random Subspaces. Those methods aim at inducing an ensemble of classifiers by producing diversity at different levels. Following this principle, Breiman has introduced in 2001 another family of methods called Random Forest. Our work aims at studying those methods in a strictly pragmatic approach, in order to provide rules on parameter settings for practitioners. For that purpose we have experimented the Forest-RI algorithm, considered as the Random Forest reference method, on the MNIST handwritten digits database. In this paper, we describe Random Forest principles and review some methods proposed in the literature. We present next our experimental protocol and results. We finally draw some conclusions on Random Forest global behavior ac- cording to their parameter tuning.

International audience; In this article, we propose a segmentation-driven recognition system which aims at extracting numerical fields from handwritten documents. We show that a crucial point of the system is the rejection ability of the handwritten numeral classifier. Therefore, we propose a simple two-stage outlier rejection strategy, and we show the benefit of this strategy on the numerical field extraction results.

International audience; In this paper, we present a method for the automatic extraction of numerical fields (ZIP codes, phone numbers, etc.) from incoming mail documents. The approach is based on a segmentation-driven recognition that aims at locating isolated and touching digits among the textual information. A syntactical analysis is then performed on each line of text in order to filter the sequences that respect a particular syntax (number of digits, presence of separators) known by the system. We evaluate the performance of our system by means of the recall precision trade-off on a real incoming mail document database.

In this paper, we show that both the writer identification and the writer verification tasks can be carried out using local features such as graphemes extracted from the segmentation of cursive handwriting. We thus enlarge the scope of the possible use of these two tasks which have been, up to now, mainly evaluated on script handwritings. A textual based Information Retrieval model is used for the writer identification stage. This allows the use of a particular feature space based on feature frequencies. Image queries are handwritten documents projected in this feature space. The approach achieves 95% correct identification on the PSI_DataBase and 86% on the IAM_DataBase. Then writer hypothesis retrieved are analysed during a verification phase. We call upon a mutual information criterion to verify that two documents may have been produced by the same writer or not. Hypothesis testing is used for this purpose. The proposed method is first scaled on the PSI_DataBase then evaluated on the IAM_DataBase. On both databases, similar performance of nearly 96% correct verification is reported, thus making the approach general and very promising for large scale applications in the domain of handwritten document querying and writer verification.

International audience; In this article, we propose a method for the automatic extraction of numerical fields in handwritten documents. The method exploits the syntax of a numerical field as an a priori knowledge to extract the connected component sequences from the document. For that, we have to label the connected components as “belonging to a numerical field” or not. We propose a method for discriminating the connected components, using different families of features and a combination of classifiers. A comparison between the results obtained with the combination of classifiers and our first approach [10] demonstrates the utility of combining different feature sets for discriminating classes with large intra-class variability.

International audience; In this paper we describe the Bovary Project, a manuscripts digitization project of the famous French writer Gustave FLAUBERT's first great work, which should end in 2006 by providing an online access to an hypertextual edition of "Madame Bovary" drafts set. We rst develop the global context of this project, the main objectives, and then focus particularly on the document analysis problem. Finally we propose a new approach for the segmentation of handwritten documents.

Recently, it has been shown that for majority voting combination methods, (negatively) dependent classifiers may provide better performance compared to that obtained with independent classifiers. The aim of this paper is to analyse the performance of plurality voting according to classifier diversity (agreement). This analysis is conducted in parallel with majority voting in order to show which method is more efficient with dependent classifiers. For this purpose, we develop a new method for the artificial generation of classifier outputs with fixed individual accuracies and pair-wise agreement. A diversity measure is applied for building the classifier teams. The experimental results show that the plurality voting is less sensitive to the correlation between classifiers than majority voting. It is also more efficient in achieving the trade- off between the recognition rate and rejection rate than the majority voting.

This thesis deals with automated image classification, applied to images acquired with alveoscopy, a new imaging technique of the distal lung. The aim is to propose and develop a computer aided-diagnosis system, so as to help the clinician analyze these images never seen before. Our contributions lie in the development of effective, robust and generic methods to classify images of healthy and pathological patients. Our first classification system is based on a rich and local characterization of the images, an ensemble of random trees approach for classification and a rejection mechanism, providing the medical expert with tools to enhance the reliability of the system. Due to the complexity of alveoscopy images and to the lack of expertize on the pathological cases (unlike healthy cases), we adopt the one-class learning paradigm which allows to learn a classifier from healthy data only. We propose a one-class approach taking advantage of combining and randomization mechanisms of ensemble methods to respond to common issues such as the curse of dimensionality. Our method is shown to be effective, robust to the dimension, competitive and even better than state-of-the-art methods on various public datasets. It has proved to be particularly relevant to our medical problem., Cette thèse porte sur la classification automatique d'images, appliquée aux images acquises par alvéoscopie, une nouvelle technique d'imagerie du poumon profond. L'objectif est la conception et le développement d'un système d'aide au diagnostic permettant d'aider le praticien à analyser ces images jamais vues auparavant. Nous avons élaboré, au travers de deux contributions, des méthodes performantes, génériques et robustes permettant de classer de façon satisfaisante les images de patients sains et pathologiques. Nous avons proposé un premier système complet de classification basé à la fois sur une caractérisation locale et riche du contenu des images, une approche de classification par méthodes d'ensemble d'arbres aléatoires et un mécanisme de pilotage du rejet de décision, fournissant à l'expert médical un moyen de renforcer la fiabilité du système. Face à la complexité des images alvéoscopiques et la difficulté de caractériser les cas pathologiques, contrairement aux cas sains, nous nous sommes orientés vers la classification one-class qui permet d'apprendre à partir des seules données des cas sains. Nous avons alors proposé une approche one-class tirant partie des mécanismes de combinaison et d'injection d'aléatoire des méthodes d'ensemble d'arbres de décision pour répondre aux difficultés rencontrées dans les approches standards, notamment la malédiction de la dimension. Les résultats obtenus montrent que notre méthode est performante, robuste à la dimension, compétitive et même meilleure comparée aux méthodes de l'état de l'art sur une grande variété de bases publiques. Elle s'est notamment avérée pertinente pour notre problématique médicale.

Despite the avenment of our world into the digital era, a large amount of handwritten documents continue to be exchanged, forcing our companies and administrations to cope with the processing of masses of documents. Automatic processing of these documents requires access to an unknown but relevant part of their content, and implies taking into account three key points: the document segmentation into relevant entities, their recognition and the rejection of irrelevant entities. Contrary to traditional approaches (full documents reading or keyword detection), all processes are parallelized leading to an information extraction approach. The first contribution of the present work is the design of a generic text line model for information extraction purpose and the implementation of a complete system based on Hidden Markov Models (HMM) constrained by this model. In one pass, the recognition module seeks to discriminate relevant information, characterized by a set of alphabetic, numeric or alphanumeric queries, with the irrelevant information, characterized by a filler model. A second contribution concerns the improvement of the local frame discrimination by using a deep neural network. This allows one to infer high-level representation for the frames and thus automate the feature extraction process. These result is a complete, generic and industrially system, responding to emerging needs in the field of handwritten document automatic reading: the extraction of complex information in unconstrained documents., Malgré l'entrée récente de notre monde dans l'ère du tout numérique, de nombreux documents manuscrits continuent à s'échanger, obligeant nos sociétés et administrations à faire face au traitement de ces masses de documents. Le traitement automatique de ces documents nécessite d'accéder à un échantillon inconnu mais pertinent de leur contenu et implique de prendre en compte trois points essentiels : la segmentation du document en entités comparable à l'information recherchée, la reconnaissance de ces entités recherchées et le rejet des entités non pertinentes. Nous nous démarquons des approches classiques de lecture complète de documents et de détection de mots clés en parallélisant ces trois traitements en une approche d'extraction d'information. Une première contribution réside dans la conception d'un modèle de ligne générique pour l'extraction d'information et l'implémentation d'un système complet à base de modèles de Markov cachés (MMC) construit autour de ce modèle. Le module de reconnaissance cherche, en une seule passe, à discriminer l'information pertinente, caractérisée par un ensemble de requêtes alphabétiques, numériques ou alphanumériques, de l'information non pertinente, caractérisée par un modèle de remplissage. Une seconde contribution réside dans l'amélioration de la discrimination locale des observations des lignes par l'utilisation d'un réseau de neurones profond. Ce dernier permet également d'inférer une représentation de haut niveau des observations et donc d'automatiser le processus d'extraction des caractéristiques. Il en résulte un système complet, générique et industrialisable, répondant à des besoins émergents dans le domaine de la lecture automatique de documents manuscrits : l'extraction d'informations complexes dans des documents non-contraints.

The automatic processing of written documents is a very active field in the industry. Indeed, due to the mass of written documents to process, the automatic analysis becomes a necessity, but the performance of current systems is highly variable according to the types of documents processed. For example, treatment of unconstrained handwritten documents remains an unresolved issue because two technological obstacles that hinder the development of reliable automatic processing of handwritten documents: - the first is the recognition of handwritten in those documents - the second is related to the existence of widely variability in the documents structures. This thesis focuses on solving the second bolt in the case of unconstrained handwritten documents. For this, we have developed reliable and robust methods to analyze document structures based on the use of Conditional Random Fields. The choice of Conditional Random Fields is motivated by the ability of these graphical models to take into account the relationships between the various entities of the document (words, phrases, blocks, ...) and integrate contextual knowledge. In addition, the use of probabilistic modeling gifted learning overcomes the inherent variability of the documents to be processed. The originality of the thesis also addresses the proposal of a hierarchical approach for extracting joint physical (segmentation of the document into blocks, lines, ...) and logical (functional interpretation of the physical structure) structures by combining low-level physical features (position, graphic, ...) and high-level logical (keyword spotting). The experiments carried out on handwritten letters show that the proposed model represents an interesting solution because of its discriminatory character and his natural ability to integrate and contextualize the characteristics of different kinds., Le traitement automatique des documents écrits est un domaine très actif dans le monde industriel. En effet, devant la masse de documents écrits à traiter, l'analyse automatique devient une nécessité mais les performances des systèmes actuels sont très variables en fonction des types de documents traités. Par exemple, le traitement des documents manuscrits non contraints reste une problématique non encore résolue à ce jour car il existe toujours deux verrous technologiques qui freinent la mise en place de systèmes fiables de traitement automatique des documents manuscrits : - la première concerne la reconnaissance des écritures manuscrites ; - la seconde est liée à l'existence d'une grande variabilité de structures de documents. Cette thèse porte sur la résolution de ce deuxième verrou dans le cas de documents manuscrits non contraints. Pour cela, nous avons développé des méthodes fiables et robustes d'analyse de structures de documents basées sur l'utilisation de Champs Aléatoires Conditionnels. Le choix des Champs Aléatoires Conditionnels est motivé par la capacité de ces modèles graphiques à prendre en compte les relations entre les différentes entités du document (mots, phrases, blocs, ...) et à intégrer des connaissances contextuelles. De plus, l'utilisation d'une modélisation probabiliste douée d'apprentissage permet de s'affranchir de la variabilité inhérente des documents à traiter. L'originalité de la thèse porte également sur la proposition d'une approche hiérarchique permettant l'extraction conjointe des structures physique (segmentation du document en blocs, lignes, ...) et logique (interprétation fonctionnelle de la structure physique) en combinant des caractéristiques physiques de bas niveau (position, représentation graphique, ...) et logiques de haut niveau (détection de mots clés). Les expérimentations effectuées sur des courriers manuscrits montrent que le modèle proposé représente une solution intéressante de par son caractère discriminant et sa capacité naturelle à intégrer et à contextualiser des caractéristiques de différentes natures.

Within the framework of the automatic processing of incoming mail documents, we present in this thesis the conception and development of a numerical field extraction system in weakly constrained handwritten documents. Although the recognition of isolated handwritten entities can be considered as a partially solved problem, the extraction of information in images of complex and free-layout documents is still a challenge. This problem requires the implementation of both handwriting recognition and information extraction methods inspired by approaches developed within the field of information extraction in electronic documents. Our contribution consists in the conception and the implementation of two different strategies: the first extends classical handwriting recognition methods, while the second is inspired from approaches used within the field of information extraction in electronic documents. The results obtained on a real handwritten mail database show that our second approach is significantly better. Finally, a complete, generic and efficient system is produced, answering one of the emergent perspectives in the field of the automatic reading of handwritten documents: the extraction of complex information in images of documents., Dans le cadre du traitement automatique de courriers entrants, nous présentons dans cette thèse l'étude, la conception et la mise en \oe uvre d'un système d'extraction de champs numériques dans des documents manuscrits quelconques. En effet, si la reconnaissance d'entités manuscrites isolées peut être considérée comme un problème en partie résolu, l'extraction d'information dans des images de documents aussi complexes et peu contraints que les courriers manuscrits libres reste à ce jour un réel défi. Ce problème nécessite aussi bien la mise en \oe uvre de méthodes classiques de reconnaissance d'entités manuscrites que de méthodes issues du domaine de l'extraction d'information dans des documents électroniques. Notre contribution repose sur le développement de deux stratégies différentes : la première réalise l'extraction des champs numériques en se basant sur les techniques classiques de reconnaissance de l'écriture, alors que la seconde, plus proche des méthodes utilisées pour l'extraction d'information, réalise indépendamment la localisation et la reconnaissance des champs. Les résultats obtenus sur une base réelle de courriers manuscrits montrent que les choix plus originaux de la seconde approche se révèlent également plus pertinents. Il en résulte un système complet, générique et industrialisable répondant à l'une des perspectives émergentes dans le domaine de la lecture automatique de documents manuscrits : l'extraction d'informations complexes dans des images de documents quelconques.