Nowadays, there are huge collections of handwritten text documents in libraries all over the world. The high demand for these resources has led to the creation of digital libraries in order to facilitate the preservation and provide electronic access to these documents. However text transcription of these documents im- ages are not always available to allow users to quickly search information, or computers to process the information, search patterns or draw out statistics. The problem is that manual transcription of these documents is an expensive task from both economical and time viewpoints. This thesis presents a novel ap- proach for e cient Computer Assisted Transcription (CAT) of handwritten text documents using state-of-the-art Handwriting Text Recognition (HTR) systems. The objective of CAT approaches is to e ciently complete a transcription task through human-machine collaboration, as the e ort required to generate a manual transcription is high, and automatically generated transcriptions from state-of-the-art systems still do not reach the accuracy required. This thesis is centered on a special application of CAT, that is, the transcription of old text document when the quantity of user e ort available is limited, and thus, the entire document cannot be revised. In this approach, the objective is to generate the best possible transcription by means of the user e ort available. This thesis provides a comprehensive view of the CAT process from feature extraction to user interaction. First, a statistical approach to generalise interactive transcription is pro- posed. As its direct application is unfeasible, some assumptions are made to apply it to two di erent tasks. First, on the interactive transcription of hand- written text documents, and next, on the interactive detection of the document layout. Next, the digitisation and annotation process of two real old text documents is described. This process was carried out because of the scarcity of similar resources and the need of annotated data to thoroughly test all the developed tools and techniques in this thesis. These two documents were carefully selected to represent the general di culties that are encountered when dealing with HTR. Baseline results are presented on these two documents to settle down a benchmark with a standard HTR system. Finally, these annotated documents were made freely available to the community. It must be noted that, all the techniques and methods developed in this thesis have been assessed on these two real old text documents. Then, a CAT approach for HTR when user e ort is limited is studied and extensively tested. The ultimate goal of applying CAT is achieved by putting together three processes. Given a recognised transcription from an HTR system. The rst process consists in locating (possibly) incorrect words and employs the user e ort available to supervise them (if necessary). As most words are not expected to be supervised due to the limited user e ort available, only a few are selected to be revised. The system presents to the user a small subset of these words according to an estimation of their correctness, or to be more precise, according to their con dence level. Next, the second process starts once these low con dence words have been supervised. This process updates the recogni- tion of the document taking user corrections into consideration, which improves the quality of those words that were not revised by the user. Finally, the last process adapts the system from the partially revised (and possibly not perfect) transcription obtained so far. In this adaptation, the system intelligently selects the correct words of the transcription. As results, the adapted system will bet- ter recognise future transcriptions. Transcription experiments using this CAT approach show that this approach is mostly e ective when user e ort is low. The last contribution of this thesis is a method for balancing the nal tran- scription quality and the supervision e ort applied using our previously de- scribed CAT approach. In other words, this method allows the user to control the amount of errors in the transcriptions obtained from a CAT approach. The motivation of this method is to let users decide on the nal quality of the desired documents, as partially erroneous transcriptions can be su cient to convey the meaning, and the user e ort required to transcribe them might be signi cantly lower when compared to obtaining a totally manual transcription. Consequently, the system estimates the minimum user e ort required to reach the amount of error de ned by the user. Error estimation is performed by computing sepa- rately the error produced by each recognised word, and thus, asking the user to only revise the ones in which most errors occur. Additionally, an interactive prototype is presented, which integrates most of the interactive techniques presented in this thesis. This prototype has been developed to be used by palaeographic expert, who do not have any background in HTR technologies. After a slight ne tuning by a HTR expert, the prototype lets the transcribers to manually annotate the document or employ the CAT ap- proach presented. All automatic operations, such as recognition, are performed in background, detaching the transcriber from the details of the system. The prototype was assessed by an expert transcriber and showed to be adequate and e cient for its purpose. The prototype is freely available under a GNU Public Licence (GPL). / Serrano Martínez-Santos, N. (2014). Interactive Transcription of Old Text Documents [Tesis doctoral no publicada]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/37979 / TESIS

[EN] Computers are now present everywhere, but their potential is not fully exploited due to some lack of acceptance. In this thesis, the pen computer paradigm is adopted, whose main idea is to replace all input devices by a pen and/or the fingers, given that the origin of the rejection comes from using unfriendly interaction devices that must be replaced by something easier for the user. This paradigm, that was was proposed several years ago, has been only recently fully implemented in products, such as the smartphones. But computers are actual illiterates that do not understand gestures or handwriting, thus a recognition step is required to "translate" the meaning of these interactions to computer-understandable language. And for this input modality to be actually usable, its recognition accuracy must be high enough. In order to realistically think about the broader deployment of pen computing, it is necessary to improve the accuracy of handwriting and gesture recognizers. This thesis is devoted to study different approaches to improve the recognition accuracy of those systems. First, we will investigate how to take advantage of interaction-derived information to improve the accuracy of the recognizer. In particular, we will focus on interactive transcription of text images. Here the system initially proposes an automatic transcript. If necessary, the user can make some corrections, implicitly validating a correct part of the transcript. Then the system must take into account this validated prefix to suggest a suitable new hypothesis. Given that in such application the user is constantly interacting with the system, it makes sense to adapt this interactive application to be used on a pen computer. User corrections will be provided by means of pen-strokes and therefore it is necessary to introduce a recognizer in charge of decoding this king of nondeterministic user feedback. However, this recognizer performance can be boosted by taking advantage of interaction-derived information, such as the user-validated prefix. Then, this thesis focuses on the study of human movements, in particular, hand movements, from a generation point of view by tapping into the kinematic theory of rapid human movements and the Sigma-Lognormal model. Understanding how the human body generates movements and, particularly understand the origin of the human movement variability, is important in the development of a recognition system. The contribution of this thesis to this topic is important, since a new technique (which improves the previous results) to extract the Sigma-lognormal model parameters is presented. Closely related to the previous work, this thesis study the benefits of using synthetic data as training. The easiest way to train a recognizer is to provide "infinite" data, representing all possible variations. In general, the more the training data, the smaller the error. But usually it is not possible to infinitely increase the size of a training set. Recruiting participants, data collection, labeling, etc., necessary for achieving this goal can be time-consuming and expensive. One way to overcome this problem is to create and use synthetically generated data that looks like the human. We study how to create these synthetic data and explore different approaches on how to use them, both for handwriting and gesture recognition. The different contributions of this thesis have obtained good results, producing several publications in international conferences and journals. Finally, three applications related to the work of this thesis are presented. First, we created Escritorie, a digital desk prototype based on the pen computer paradigm for transcribing handwritten text images. Second, we developed "Gestures à Go Go", a web application for bootstrapping gestures. Finally, we studied another interactive application under the pen computer paradigm. In this case, we study how translation reviewing can be done more ergonomically using a pen. / [ES] Hoy en día, los ordenadores están presentes en todas partes pero su potencial no se aprovecha debido al "miedo" que se les tiene. En esta tesis se adopta el paradigma del pen computer, cuya idea fundamental es sustituir todos los dispositivos de entrada por un lápiz electrónico o, directamente, por los dedos. El origen del rechazo a los ordenadores proviene del uso de interfaces poco amigables para el humano. El origen de este paradigma data de hace más de 40 años, pero solo recientemente se ha comenzado a implementar en dispositivos móviles. La lenta y tardía implantación probablemente se deba a que es necesario incluir un reconocedor que "traduzca" los trazos del usuario (texto manuscrito o gestos) a algo entendible por el ordenador. Para pensar de forma realista en la implantación del pen computer, es necesario mejorar la precisión del reconocimiento de texto y gestos. El objetivo de esta tesis es el estudio de diferentes estrategias para mejorar esta precisión. En primer lugar, esta tesis investiga como aprovechar información derivada de la interacción para mejorar el reconocimiento, en concreto, en la transcripción interactiva de imágenes con texto manuscrito. En la transcripción interactiva, el sistema y el usuario trabajan "codo con codo" para generar la transcripción. El usuario valida la salida del sistema proporcionando ciertas correcciones, mediante texto manuscrito, que el sistema debe tener en cuenta para proporcionar una mejor transcripción. Este texto manuscrito debe ser reconocido para ser utilizado. En esta tesis se propone aprovechar información contextual, como por ejemplo, el prefijo validado por el usuario, para mejorar la calidad del reconocimiento de la interacción. Tras esto, la tesis se centra en el estudio del movimiento humano, en particular del movimiento de las manos, utilizando la Teoría Cinemática y su modelo Sigma-Lognormal. Entender como se mueven las manos al escribir, y en particular, entender el origen de la variabilidad de la escritura, es importante para el desarrollo de un sistema de reconocimiento, La contribución de esta tesis a este tópico es importante, dado que se presenta una nueva técnica (que mejora los resultados previos) para extraer el modelo Sigma-Lognormal de trazos manuscritos. De forma muy relacionada con el trabajo anterior, se estudia el beneficio de utilizar datos sintéticos como entrenamiento. La forma más fácil de entrenar un reconocedor es proporcionar un conjunto de datos "infinito" que representen todas las posibles variaciones. En general, cuanto más datos de entrenamiento, menor será el error del reconocedor. No obstante, muchas veces no es posible proporcionar más datos, o hacerlo es muy caro. Por ello, se ha estudiado como crear y usar datos sintéticos que se parezcan a los reales. Las diferentes contribuciones de esta tesis han obtenido buenos resultados, produciendo varias publicaciones en conferencias internacionales y revistas. Finalmente, también se han explorado tres aplicaciones relaciones con el trabajo de esta tesis. En primer lugar, se ha creado Escritorie, un prototipo de mesa digital basada en el paradigma del pen computer para realizar transcripción interactiva de documentos manuscritos. En segundo lugar, se ha desarrollado "Gestures à Go Go", una aplicación web para generar datos sintéticos y empaquetarlos con un reconocedor de forma rápida y sencilla. Por último, se presenta un sistema interactivo real bajo el paradigma del pen computer. En este caso, se estudia como la revisión de traducciones automáticas se puede realizar de forma más ergonómica. / [CAT] Avui en dia, els ordinadors són presents a tot arreu i es comunament acceptat que la seva utilització proporciona beneficis. No obstant això, moltes vegades el seu potencial no s'aprofita totalment. En aquesta tesi s'adopta el paradigma del pen computer, on la idea fonamental és substituir tots els dispositius d'entrada per un llapis electrònic, o, directament, pels dits. Aquest paradigma postula que l'origen del rebuig als ordinadors prové de l'ús d'interfícies poc amigables per a l'humà, que han de ser substituïdes per alguna cosa més coneguda. Per tant, la interacció amb l'ordinador sota aquest paradigma es realitza per mitjà de text manuscrit i/o gestos. L'origen d'aquest paradigma data de fa més de 40 anys, però només recentment s'ha començat a implementar en dispositius mòbils. La lenta i tardana implantació probablement es degui al fet que és necessari incloure un reconeixedor que "tradueixi" els traços de l'usuari (text manuscrit o gestos) a alguna cosa comprensible per l'ordinador, i el resultat d'aquest reconeixement, actualment, és lluny de ser òptim. Per pensar de forma realista en la implantació del pen computer, cal millorar la precisió del reconeixement de text i gestos. L'objectiu d'aquesta tesi és l'estudi de diferents estratègies per millorar aquesta precisió. En primer lloc, aquesta tesi investiga com aprofitar informació derivada de la interacció per millorar el reconeixement, en concret, en la transcripció interactiva d'imatges amb text manuscrit. En la transcripció interactiva, el sistema i l'usuari treballen "braç a braç" per generar la transcripció. L'usuari valida la sortida del sistema donant certes correccions, que el sistema ha d'usar per millorar la transcripció. En aquesta tesi es proposa utilitzar correccions manuscrites, que el sistema ha de reconèixer primer. La qualitat del reconeixement d'aquesta interacció és millorada, tenint en compte informació contextual, com per exemple, el prefix validat per l'usuari. Després d'això, la tesi se centra en l'estudi del moviment humà en particular del moviment de les mans, des del punt de vista generatiu, utilitzant la Teoria Cinemàtica i el model Sigma-Lognormal. Entendre com es mouen les mans en escriure és important per al desenvolupament d'un sistema de reconeixement, en particular, per entendre l'origen de la variabilitat de l'escriptura. La contribució d'aquesta tesi a aquest tòpic és important, atès que es presenta una nova tècnica (que millora els resultats previs) per extreure el model Sigma- Lognormal de traços manuscrits. De forma molt relacionada amb el treball anterior, s'estudia el benefici d'utilitzar dades sintètiques per a l'entrenament. La forma més fàcil d'entrenar un reconeixedor és proporcionar un conjunt de dades "infinit" que representin totes les possibles variacions. En general, com més dades d'entrenament, menor serà l'error del reconeixedor. No obstant això, moltes vegades no és possible proporcionar més dades, o fer-ho és molt car. Per això, s'ha estudiat com crear i utilitzar dades sintètiques que s'assemblin a les reals. Les diferents contribucions d'aquesta tesi han obtingut bons resultats, produint diverses publicacions en conferències internacionals i revistes. Finalment, també s'han explorat tres aplicacions relacionades amb el treball d'aquesta tesi. En primer lloc, s'ha creat Escritorie, un prototip de taula digital basada en el paradigma del pen computer per realitzar transcripció interactiva de documents manuscrits. En segon lloc, s'ha desenvolupat "Gestures à Go Go", una aplicació web per a generar dades sintètiques i empaquetar-les amb un reconeixedor de forma ràpida i senzilla. Finalment, es presenta un altre sistema inter- actiu sota el paradigma del pen computer. En aquest cas, s'estudia com la revisió de traduccions automàtiques es pot realitzar de forma més ergonòmica. / Martín-Albo Simón, D. (2016). Contributions to Pen & Touch Human-Computer Interaction [Tesis doctoral no publicada]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/68482 / TESIS

Nowadays, the main libraries and document archives are investing a considerable effort on digitizing their collections. Indeed, most of them are scanning the documents and publishing the resulting images without their corresponding transcriptions. This seriously limits the document exploitation possibilities. When the transcription is necessary, it is manually performed by human experts, which is a very expensive and error-prone task. Obtaining transcriptions to the level of required quality demands the intervention of human experts to review and correct the resulting output of the recognition engines. To this end, it is extremely useful to provide interactive tools to obtain and edit the transcription. Although text recognition is the final goal, several previous steps (known as preprocessing) are necessary in order to get a fine transcription from a digitized image. Document cleaning, enhancement, and binarization (if they are needed) are the first stages of the recognition pipeline. Historical Handwritten Documents, in addition, show several degradations, stains, ink-trough and other artifacts. Therefore, more sophisticated and elaborate methods are required when dealing with these kind of documents, even expert supervision in some cases is needed. Once images have been cleaned, main zones of the image have to be detected: those that contain text and other parts such as images, decorations, versal letters. Moreover, the relations among them and the final text have to be detected. Those preprocessing steps are critical for the final performance of the system since an error at this point will be propagated during the rest of the transcription process. The ultimate goal of the Document Image Analysis pipeline is to receive the transcription of the text (Optical Character Recognition and Handwritten Text Recognition). During this thesis we aimed to improve the main stages of the recognition pipeline, from the scanned documents as input to the final transcription. We focused our effort on applying Neural Networks and deep learning techniques directly on the document images to extract suitable features that will be used by the different tasks dealt during the following work: Image Cleaning and Enhancement (Document Image Binarization), Layout Extraction, Text Line Extraction, Text Line Normalization and finally decoding (or text line recognition). As one can see, the following work focuses on small improvements through the several Document Image Analysis stages, but also deals with some of the real challenges: historical manuscripts and documents without clear layouts or very degraded documents. Neural Networks are a central topic for the whole work collected in this document. Different convolutional models have been applied for document image cleaning and enhancement. Connectionist models have been used, as well, for text line extraction: first, for detecting interest points and combining them in text segments and, finally, extracting the lines by means of aggregation techniques; and second, for pixel labeling to extract the main body area of the text and then the limits of the lines. For text line preprocessing, i.e., to normalize the text lines before recognizing them, similar models have been used to detect the main body area and then to height-normalize the images giving more importance to the central area of the text. Finally, Convolutional Neural Networks and deep multilayer perceptrons have been combined with hidden Markov models to improve our transcription engine significantly. The suitability of all these approaches has been tested with different corpora for any of the stages dealt, giving competitive results for most of the methodologies presented. / Hoy en día, las principales librerías y archivos está invirtiendo un esfuerzo considerable en la digitalización de sus colecciones. De hecho, la mayoría están escaneando estos documentos y publicando únicamente las imágenes sin transcripciones, limitando seriamente la posibilidad de explotar estos documentos. Cuando la transcripción es necesaria, esta se realiza normalmente por expertos de forma manual, lo cual es una tarea costosa y propensa a errores. Si se utilizan sistemas de reconocimiento automático se necesita la intervención de expertos humanos para revisar y corregir la salida de estos motores de reconocimiento. Por ello, es extremadamente útil para proporcionar herramientas interactivas con el fin de generar y corregir la transcripciones. Aunque el reconocimiento de texto es el objetivo final del Análisis de Documentos, varios pasos previos (preprocesamiento) son necesarios para conseguir una buena transcripción a partir de una imagen digitalizada. La limpieza, mejora y binarización de las imágenes son las primeras etapas del proceso de reconocimiento. Además, los manuscritos históricos tienen una mayor dificultad en el preprocesamiento, puesto que pueden mostrar varios tipos de degradaciones, manchas, tinta a través del papel y demás dificultades. Por lo tanto, este tipo de documentos requiere métodos de preprocesamiento más sofisticados. En algunos casos, incluso, se precisa de la supervisión de expertos para garantizar buenos resultados en esta etapa. Una vez que las imágenes han sido limpiadas, las diferentes zonas de la imagen deben de ser localizadas: texto, gráficos, dibujos, decoraciones, letras versales, etc. Por otra parte, también es importante conocer las relaciones entre estas entidades. Estas etapas del pre-procesamiento son críticas para el rendimiento final del sistema, ya que los errores cometidos en aquí se propagarán al resto del proceso de transcripción. El objetivo principal del trabajo presentado en este documento es mejorar las principales etapas del proceso de reconocimiento completo: desde las imágenes escaneadas hasta la transcripción final. Nuestros esfuerzos se centran en aplicar técnicas de Redes Neuronales (ANNs) y aprendizaje profundo directamente sobre las imágenes de los documentos, con la intención de extraer características adecuadas para las diferentes tareas: Limpieza y Mejora de Documentos, Extracción de Líneas, Normalización de Líneas de Texto y, finalmente, transcripción del texto. Como se puede apreciar, el trabajo se centra en pequeñas mejoras en diferentes etapas del Análisis y Procesamiento de Documentos, pero también trata de abordar tareas más complejas: manuscritos históricos, o documentos que presentan degradaciones. Las ANNs y el aprendizaje profundo son uno de los temas centrales de esta tesis. Diferentes modelos neuronales convolucionales se han desarrollado para la limpieza y mejora de imágenes de documentos. También se han utilizado modelos conexionistas para la extracción de líneas: primero, para detectar puntos de interés y segmentos de texto y, agregarlos para extraer las líneas del documento; y en segundo lugar, etiquetando directamente los píxeles de la imagen para extraer la zona central del texto y así definir los límites de las líneas. Para el preproceso de las líneas de texto, es decir, la normalización del texto antes del reconocimiento final, se han utilizado modelos similares a los mencionados para detectar la zona central del texto. Las imagenes se rescalan a una altura fija dando más importancia a esta zona central. Por último, en cuanto a reconocimiento de escritura manuscrita, se han combinado técnicas de ANNs y aprendizaje profundo con Modelos Ocultos de Markov, mejorando significativamente los resultados obtenidos previamente por nuestro motor de reconocimiento. La idoneidad de todos estos enfoques han sido testeados con diferentes corpus en cada una de las tareas tratadas., obtenie / Avui en dia, les principals llibreries i arxius històrics estan invertint un esforç considerable en la digitalització de les seues col·leccions de documents. De fet, la majoria estan escanejant aquests documents i publicant únicament les imatges sense les seues transcripcions, fet que limita seriosament la possibilitat d'explotació d'aquests documents. Quan la transcripció del text és necessària, normalment aquesta és realitzada per experts de forma manual, la qual cosa és una tasca costosa i pot provocar errors. Si s'utilitzen sistemes de reconeixement automàtic es necessita la intervenció d'experts humans per a revisar i corregir l'eixida d'aquests motors de reconeixement. Per aquest motiu, és extremadament útil proporcionar eines interactives amb la finalitat de generar i corregir les transcripcions generades pels motors de reconeixement. Tot i que el reconeixement del text és l'objectiu final de l'Anàlisi de Documents, diversos passos previs (coneguts com preprocessament) són necessaris per a l'obtenció de transcripcions acurades a partir d'imatges digitalitzades. La neteja, millora i binarització de les imatges (si calen) són les primeres etapes prèvies al reconeixement. A més a més, els manuscrits històrics presenten una major dificultat d'analisi i preprocessament, perquè poden mostrar diversos tipus de degradacions, taques, tinta a través del paper i altres peculiaritats. Per tant, aquest tipus de documents requereixen mètodes de preprocessament més sofisticats. En alguns casos, fins i tot, es precisa de la supervisió d'experts per a garantir bons resultats en aquesta etapa. Una vegada que les imatges han sigut netejades, les diferents zones de la imatge han de ser localitzades: text, gràfics, dibuixos, decoracions, versals, etc. D'altra banda, també és important conéixer les relacions entre aquestes entitats i el text que contenen. Aquestes etapes del preprocessament són crítiques per al rendiment final del sistema, ja que els errors comesos en aquest moment es propagaran a la resta del procés de transcripció. L'objectiu principal del treball que estem presentant és millorar les principals etapes del procés de reconeixement, és a dir, des de les imatges escanejades fins a l'obtenció final de la transcripció del text. Els nostres esforços se centren en aplicar tècniques de Xarxes Neuronals (ANNs) i aprenentatge profund directament sobre les imatges de documents, amb la intenció d'extraure característiques adequades per a les diferents tasques analitzades: neteja i millora de documents, extracció de línies, normalització de línies de text i, finalment, transcripció. Com es pot apreciar, el treball realitzat aplica xicotetes millores en diferents etapes de l'Anàlisi de Documents, però també tracta d'abordar tasques més complexes: manuscrits històrics, o documents que presenten degradacions. Les ANNs i l'aprenentatge profund són un dels temes centrals d'aquesta tesi. Diferents models neuronals convolucionals s'han desenvolupat per a la neteja i millora de les dels documents. També s'han utilitzat models connexionistes per a la tasca d'extracció de línies: primer, per a detectar punts d'interés i segments de text i, agregar-los per a extraure les línies del document; i en segon lloc, etiquetant directament els pixels de la imatge per a extraure la zona central del text i així definir els límits de les línies. Per al preprocés de les línies de text, és a dir, la normalització del text abans del reconeixement final, s'han utilitzat models similars als utilitzats per a l'extracció de línies. Finalment, quant al reconeixement d'escriptura manuscrita, s'han combinat tècniques de ANNs i aprenentatge profund amb Models Ocults de Markov, que han millorat significativament els resultats obtinguts prèviament pel nostre motor de reconeixement. La idoneïtat de tots aquests enfocaments han sigut testejats amb diferents corpus en cadascuna de les tasques tractad / Pastor Pellicer, J. (2017). Neural Networks for Document Image and Text Processing [Tesis doctoral no publicada]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/90443 / TESIS

Cette thèse porte sur le développement d'une chaîne de traitement complète pour réaliser des tâches de reconnaissance d'écriture manuscrite non contrainte. Trois difficultés majeures sont à résoudre: l'étape du prétraitement, l'étape de la modélisation optique et l'étape de la modélisation du langage. Au stade des prétraitements il faut extraire correctement les lignes de texte à partir de l'image du document. Une méthode de segmentation itérative en lignes utilisant des filtres orientables a été développée à cette fin. La difficulté dans l’étape de la modélisation optique vient de la diversité stylistique des scripts d'écriture manuscrite. Les modèles optiques statistiques développés sont des modèles de Markov cachés (HMM-GMM) et les modèles de réseaux de neurones récurrents (BLSTM-CTC). Les réseaux récurrents permettent d’atteindre les performances de l’état de l’art sur les deux bases de référence RIMES (pour le Français) et IAM (pour l’anglais). L'étape de modélisation du langage implique l'intégration d’un lexique et d’un modèle de langage statistique afin de rechercher parmi les hypothèses proposées par le modèle optique, la séquence de mots (phrase) la plus probable du point de vue linguistique. La difficulté à ce stade est liée à l’obtention d’un modèle de couverture lexicale optimale avec un minimum de mots hors vocabulaire (OOV). Pour cela nous introduisons une modélisation en sous-unités lexicales composée soit de syllabes soit de multigrammes. Ces modèles couvrent efficacement une partie importante des mots hors vocabulaire. Les performances du système de reconnaissance avec les unités sous-lexicales dépassent les performances des systèmes de reconnaissance traditionnelles de mots ou de caractères en présence d’un fort taux de mots hors lexique. Elles sont équivalentes aux modèles traditionnels en présence d’un faible taux de mots hors lexique. Grâce à la taille compacte du modèle de langage reposant sur des unités sous-lexicales, un système de reconnaissance multilingue unifié a été réalisé. Le système multilingue unifié améliore les performances de reconnaissance par rapport aux systèmes spécialisés dans chaque langue, notamment lorsque le modèle optique unifié est utilisé. / This thesis is about the design of a complete processing chain dedicated to unconstrained handwriting recognition. Three main difficulties are adressed: pre-processing, optical modeling and language modeling. The pre-processing stage is related to extracting properly the text lines to be recognized from the document image. An iterative text line segmentation method using oriented steerable filters was developed for this purpose. The difficulty in the optical modeling stage lies in style diversity of the handwriting scripts. Statistical optical models are traditionally used to tackle this problem such as Hidden Markov models (HMM-GMM) and more recently recurrent neural networks (BLSTM-CTC). Using BLSTM we achieve state of the art performance on the RIMES (for French) and IAM (for English) datasets. The language modeling stage implies the integration of a lexicon and a statistical language model to the recognition processing chain in order to constrain the recognition hypotheses to the most probable sequence of words (sentence) from the language point of view. The difficulty at this stage is related to the finding the optimal vocabulary with minimum Out-Of-Vocabulary words rate (OOV). Enhanced language modeling approaches has been introduced by using sub-lexical units made of syllables or multigrams. The sub-lexical units cover an important portion of the OOV words. Then the language coverage depends on the domain of the language model training corpus, thus the need to train the language model with in domain data. The recognition system performance with the sub-lexical units outperformes the traditional recognition systems that use words or characters language models, in case of high OOV rates. Otherwise equivalent performances are obtained with a compact sub-lexical language model. Thanks to the compact lexicon size of the sub-lexical units, a unified multilingual recognition system has been designed. The unified system performance have been evaluated on the RIMES and IAM datasets. The unified multilingual system shows enhanced recognition performance over the specialized systems, especially when a unified optical model is used.

Avec l’avènement des dispositifs nomades tels que les smartphones et les tablettes, la reconnaissance automatique de l’écriture manuscrite cursive à partir d’un signal en ligne est devenue durant les dernières décennies un besoin réel de la vie quotidienne à l’ère numérique. Dans le cadre de cette thèse, nous proposons de nouvelles stratégies pour un système de reconnaissance de mots manuscrits en-ligne. Ce système se base sur une méthode collaborative segmentation/reconnaissance et en utilisant des analyses à deux niveaux : caractère et bi-caractères. Plus précisément, notre système repose sur une segmentation de mots manuscrits en graphèmes afin de créer un treillis à L niveaux. Chaque noeud de ce treillis est considéré comme un caractère potentiel envoyé à un moteur de Reconnaissance de Caractères Isolés (RCI) basé sur un SVM. Pour chaque noeud, ce dernier renvoie une liste de caractères associés à une liste d’estimations de probabilités de reconnaissance. Du fait de la grande diversité des informations résultant de la segmentation en graphèmes, en particulier à cause de la présence de morceaux de caractères et de ligatures, l’injection de chacun des noeuds du treillis dans le RCI engendre de potentielles ambiguïtés au niveau du caractère. Nous proposons de lever ces ambiguïtés en utilisant des modèles de bi-caractères, basés sur une régression logistique dont l’objectif est de vérifier la cohérence des informations à un niveau de reconnaissance plus élevé. Finalement, les résultats renvoyés par le RCI et l’analyse des modèles de bi-caractères sont utilisés dans la phase de décodage pour parcourir le treillis dans le but de trouver le chemin optimal associé à chaque mot dans le lexique. Deux méthodes de décodage sont proposées (recherche heuristique et programmation dynamique), la plus efficace étant basée sur de la programmation dynamique. / With the advent of mobile devices such as tablets and smartphones over the last decades, on-line handwriting recognition has become a very highly demanded service for daily life activities and professional applications. This thesis presents a new approach for on-line handwriting recognition. This approach is based on explicit segmentation/recognition integrated in a two level analysis system: character and bi-character. More specifically, our system segments a handwritten word in a sequence of graphemes to be then used to create a L-levels lattice of graphemes. Each node of the lattice is considered as a character to be submitted to a SVM based Isolated Character Recognizer (ICR). The ICR returns a list of potential character candidates, each of which is associated with an estimated recognition probability. However, each node of the lattice is a combination of various segmented graphemes. As a consequence, a node may contain some ambiguous information that cannot be handled by the ICR at character level analysis. We propose to solve this problem using "bi-character" models based on Logistic Regression, in order to verify the consistency of the information at a higher level of analysis. Finally, the recognition results provided by the ICR and the bi-character models are used in the word decoding stage, whose role is to find the optimal path in the lattice associated to each word in the lexicon. Two methods are presented for word decoding (heuristic search and dynamic programming), and dynamic programming is found to be the most effective.

We use a Fully Convolutional Neural Network (FCNN) to classify pixels in historical document images, enabling the extraction of high-quality, pixel-precise and semantically consistent layers of masked content. We also analyze a dataset of hand-labeled historical form images of unprecedented detail and complexity. The semantic categories we consider in this new dataset include handwriting, machine-printed text, dotted and solid lines, and stamps. Segmentation of document images into distinct layers allows handwriting, machine print, and other content to be processed and recognized discriminatively, and therefore more intelligently than might be possible with content-unaware methods. We show that an efficient FCNN with relatively few parameters can accurately segment documents having similar textural content when trained on a single representative pixel-labeled document image, even when layouts differ significantly. In contrast to the overwhelming majority of existing semantic segmentation approaches, we allow multiple labels to be predicted per pixel location, which allows for direct prediction and reconstruction of overlapped content. We perform an analysis of prevalent pixel-wise performance measures, and show that several popular performance measures can be manipulated adversarially, yielding arbitrarily high measures based on the type of bias used to generate the ground-truth. We propose a solution to the gaming problem by comparing absolute performance to an estimated human level of performance. We also present results on a recent international competition requiring the automatic annotation of billions of pixels, in which our method took first place.

We use a Fully Convolutional Neural Network (FCNN) to classify pixels in historical document images, enabling the extraction of high-quality, pixel-precise and semantically consistent layers of masked content. We also analyze a dataset of hand-labeled historical form images of unprecedented detail and complexity. The semantic categories we consider in this new dataset include handwriting, machine-printed text, dotted and solid lines, and stamps. Segmentation of document images into distinct layers allows handwriting, machine print, and other content to be processed and recognized discriminatively, and therefore more intelligently than might be possible with content-unaware methods. We show that an efficient FCNN with relatively few parameters can accurately segment documents having similar textural content when trained on a single representative pixel-labeled document image, even when layouts differ significantly. In contrast to the overwhelming majority of existing semantic segmentation approaches, we allow multiple labels to be predicted per pixel location, which allows for direct prediction and reconstruction of overlapped content. We perform an analysis of prevalent pixel-wise performance measures, and show that several popular performance measures can be manipulated adversarially, yielding arbitrarily high measures based on the type of bias used to generate the ground-truth. We propose a solution to the gaming problem by comparing absolute performance to an estimated human level of performance. We also present results on a recent international competition requiring the automatic annotation of billions of pixels, in which our method took first place.

Les méthodes à l’état de l’art de la reconnaissance de l’écriture sont fondées sur des réseaux de neurones récurrents (RNN) à cellules LSTM ayant des performances remarquables. Dans cette thèse, nous proposons deux nouveaux principes la vérification lexicale et la génération de cohorte afin d’attaquer les problèmes de la reconnaissance de l’écriture : i) le problème des grands lexiques et des décodages dirigés par le lexique ii) la problématique de combinaison de modèles optiques pour une meilleure reconnaissance iii) la nécessité de constituer de très grands ensembles de données étiquetées dans un contexte d’apprentissage profond. La vérification lexicale est une alternative aux décodages dirigés par le lexique peu étudiée à cause des faibles performances des modèles optiques historiques (HMM). Nous montrons dans cette thèse qu’elle constitue une alternative intéressante aux approches dirigées par le lexique lorsqu’elles s’appuient sur des modèles optiques très performants comme les RNN LSTM. La génération de cohorte permet de générer facilement et rapidement un grand nombre de réseaux récurrents complémentaires en un seul apprentissage. De ces deux techniques nous construisons et proposons un nouveau schéma de cascade pour la reconnaissance de mots isolés, une nouvelle combinaison au niveau ligne LV-ROVER et une nouvelle stratégie d’auto-apprentissage de RNN LSTM pour la reconnaissance de mots isolés. La cascade proposée permet de combiner avec la vérification lexicale des milliers de réseaux et atteint des résultats à l’état de l’art pour les bases Rimes et IAM. LV-ROVER a une complexité réduite par rapport à l’algorithme original ROVER et permet de combiner des centaines de réseaux sans modèle de langage tout en dépassant l’état de l’art pour la reconnaissance de lignes sur le jeu de donnéesRimes. Notre stratégie d’auto-apprentissage permet d’apprendre à partir d’un seul réseau BLSTM et sans paramètres grâce à la cohorte et la vérification lexicale, elle montre d’excellents résultats sur les bases Rimes et IAM. / State-of-the-art methods for handwriting recognition are based on LSTM recurrent neural networks (RNN) which achieve high performance recognition. In this thesis, we propose the lexicon verification and the cohort generation as two new building blocs to tackle the problem of handwriting recognition which are : i) the large vocabulary problem and the use of lexicon driven methods ii) the combination of multiple optical models iii) the need for large labeled dataset for training RNN. The lexicon verification is an alternative to the lexicon driven decoding process and can deal with lexicons of 3 millions words. The cohort generation is a method to get easily and quickly a large number of complementary recurrent neural networks extracted from a single training. From these two new techniques we build and propose a new cascade scheme for isolated word recognition, a new line level combination LV-ROVER and a new self-training strategy to train LSTM RNN for isolated handwritten words recognition. The proposed cascade combines thousands of LSTM RNN with lexicon verification and achieves state-of-the art word recognition performance on the Rimes and IAM datasets. The Lexicon Verified ROVER : LV-ROVER, has a reduce complexity compare to the original ROVER algorithm and combine hundreds of recognizers without language models while achieving state of the art for handwritten line text on the RIMES dataset. Our self-training strategy use both labeled and unlabeled data with the unlabeled data being self-labeled by its own lexicon verified predictions. The strategy enables self-training with a single BLSTM and show excellent results on the Rimes and Iam datasets.

No description available.

In this thesis, we address some of the challenges involved in developing a robust writer-independent, lexicon-free system to recognize online Tamil words. Tamil, being a Dravidian language, is morphologically rich and also agglutinative and thus does not have a finite lexicon. For example, a single verb root can easily lead to hundreds of words after morphological changes and agglutination. Further, adoption of a lexicon-free recognition approach can be applied to form-filling applications, wherein the lexicon can become cumbersome (if not impossible) to capture all possible names. Under such circumstances, one must necessarily explore the possibility of segmenting a Tamil word to its individual symbols. Modern day Tamil alphabet comprises 23 consonants and 11 vowels forming a total combination of 313 characters/aksharas. A minimal set of 155 distinct symbols have been derived to recognize these characters. A corpus of isolated Tamil symbols (IWFHR database) is used for deriving the various statistics proposed in this work. To address the challenges of segmentation and recognition (the primary focus of the thesis), Tamil words are collected using a custom application running on a tablet PC. A set of 10000 words (comprising 53246 symbols) have been collected from high school students and used for the experiments in this thesis. We refer to this database as the ‘MILE word database’. In the first part of the work, a feedback based word segmentation mechanism has been proposed. Initially, the Tamil word is segmented based on a bounding box overlap criterion. This dominant overlap criterion segmentation (DOCS) generates a set of candidate stroke groups. Thereafter, attention is paid to certain attributes from the resulting stroke groups for detecting any possible splits or under-segmentations. By relying on feedbacks provided by a priori knowledge of attributes such as number of dominant points and inter-stroke displacements the recognition label and likelihood of the primary SVM classifier linguistic knowledge on the detected stroke groups, a decision is taken to correct it or not. Accordingly, we call the proposed segmentation as ‘attention feedback segmentation’ (AFS). Across the words in the MILE word database, a segmentation rate of 99.7% is achieved at symbol level with AFS. The high segmentation rate (with feedback) in turn improves the symbol recognition rate of the primary SVM classifier from 83.9% (with DOCS alone) to 88.4%. For addressing the problem of segmentation, the SVM classifier fed with the x-y trace of the normalized and resampled online stroke groups is quite effective. However, the performance of the classifier is not robust to effectively distinguish between many sets of similar looking symbols. In order to improve the symbol recognition performance, we explore two approaches, namely reevaluation strategies and language models. The reevaluation techniques, in particular, resolve the ambiguities in base consonants, pure consonants and vowel modifiers to a considerable extent. For the frequently confused sets (derived from the confusion matrix), a dynamic time warping (DTW) approach is proposed to automatically extract their discriminative regions. Dedicated to each confusion set, novel localized cues are derived from the discriminative region for their disambiguation. The proposed features are quite promising in improving the symbol recognition performance of the confusion sets. Comparative experimental analysis of these features with x-y coordinates are performed for judging their discriminative power. The resolving of confusions is accomplished with expert networks, comprising discriminative region extractor, feature extractor and SVM. The proposed techniques improve the symbol recognition rate by 3.5% (from 88.4% to 91.9%) on the MILE word database over the primary SVM classifier. In the final part of the thesis, we integrate linguistic knowledge (derived from a text corpus) in the primary recognition system. The biclass, bigram and unigram language models at symbol level are compared in terms of recognition performance. Amongst the three models, the bigram model is shown to give the highest recognition accuracy. A class reduction approach for recognition is adopted by incorporating the language bigram model at the akshara level. Lastly, a judicious combination of reevaluation techniques with language models is proposed in this work. Overall, an improvement of up to 4.7% (from 88.4% to 93.1%) in symbol level accuracy is achieved. The writer-independent and lexicon-free segmentation-recognition approach developed in this thesis for online handwritten Tamil word recognition is promising. The best performance of 93.1% (achieved at symbol level) is comparable to the highest reported accuracy in the literature for Tamil symbols. However, the latter one is on a database of isolated symbols (IWFHR competition test dataset), whereas our accuracy is on a database of 10000 words and thus, a product of segmentation and classifier accuracies. The recognition performance obtained may be enhanced further by experimenting on and choosing the best set of features and classifiers. Also, the word recognition performance can be very significantly improved by using a lexicon. However, these are not the issues addressed by the thesis. We hope that the lexicon-free experiments reported in this work will serve as a benchmark for future efforts.