<p>The computer vision task of classifying natural images is a primary driving force behind modern AI algorithms. Deep Convolutional Neural Networks (CNNs) demonstrate state of the art performance in large scale multi-class image classification tasks. However, due to the many layers and millions of parameters these models are considered to be black box algorithms. The decisions of these models are further obscured due to a cumbersome multi-class decision process. There exists another approach called class binarization in the literature which determines the multi-class prediction outcome through a sequence of binary decisions.The focus of this dissertation is on the integration of the class-binarization approach to multi-class classification with deep learning models, such as CNNs, for addressing large scale image classification problems. Three works are presented to address the integration.</p> <p>In the first work, Error Correcting Output Codes (ECOCs) are integrated into CNNs by inserting a latent-binarization layer prior to the CNNs final classification layer.  This approach encapsulates both encoding and decoding steps of ECOC into a single CNN architecture. EM and Gibbs sampling algorithms are combined with back-propagation to train CNN models with Latent Binarization (LB-CNN). The training process of LB-CNN guides the model to discover hidden relationships similar to the semantic relationships known apriori between the categories. The proposed models and algorithms are applied to several image recognition tasks, producing excellent results.</p> <p>In the second work, Hierarchically Decodeable Output Codes (HD-OCs) are proposedto compactly describe a hierarchical probabilistic binary decision process model over the features of a CNN. HD-OCs enforce more homogeneous assignments of the categories to the dichotomy labels. A novel concept called average decision depth is presented to quantify the average number of binary questions needed to classify an input. An HD-OC is trained using a hierarchical log-likelihood loss that is empirically shown to orient the output of the latent feature space to resemble the hierarchical structure described by the HD-OC. Experiments are conducted at several different scales of category labels. The experiments demonstrate strong performance and powerful insights into the decision process of the model.</p> <p>In the final work, the literature of enumerative combinatorics and partially ordered sets isused to establish a unifying framework of class-binarization methods under the Multivariate Bernoulli family of models. The unifying framework theoretically establishes simple relationships for transitioning between the different binarization approaches. Such relationships provide useful investigative tools for the discovery of statistical dependencies between large groups of categories. They are additionally useful for incorporating taxonomic information as well as enforcing structural model constraints. The unifying framework lays the groundwork for future theoretical and methodological work in addressing the fundamental issues of large scale multi-class classification.</p> <p><br></p>

Text is no longer confined to scanned pages and often appears in camera-based images originating from text on real world objects. Unlike the images from conventional flatbed scanners, which have a controlled acquisition environment, camera-based images pose new challenges such as uneven illumination, blur, poor resolution, perspective distortion and 3D deformations that can severely affect the performance of any optical character recognition (OCR) system. Due to the variations in the imaging condition as well as the target document type, traditional OCR systems, designed for scanned images, cannot be directly applied to camera-captured images and a new level of processing needs to be addressed. In this thesis, we study some of the issues commonly encountered in camera-based image analysis and propose novel methods to overcome them. All the methods make use of color connected components. 1. Connected component descriptor for document image mosaicing Document image analysis often requires mosaicing when it is not possible to capture a large document at a reasonable resolution in a single exposure. Such a document is captured in parts and mosaicing stitches them into a single image. Since connected components (CCs) in a document image can easily be extracted regardless of the image rotation, scale and perspective distortion, we design a robust feature named connected component descriptor that is tailored for mosaicing camera-captured document images. The method involves extraction of a circular measurement region around each CC and its description using the angular radial transform (ART). To ensure geometric consistency during feature matching, the ART coefficients of a CC are augmented with those of its 2 nearest neighbors. Our method addresses two critical issues often encountered in correspondence matching: (i) the stability of features and (ii) robustness against false matches due to multiple instances of many characters in a document image. We illustrate the effectiveness of the proposed method on camera-captured document images exhibiting large variations in viewpoint, illumination and scale. 2. Font and background color independent text binarization The first step in an OCR system, after document acquisition, is binarization, which converts a gray-scale/color image into a two-level image -the foreground text and the background. We propose two methods for binarization of color documents whereby the foreground text is output as black and the background as white regardless of the polarity of foreground-background shades. (a) Hierarchical CC Analysis: The method employs an edge-based connected component approach and automatically determines a threshold for each component. It overcomes several limitations of existing locally-adaptive thresholding techniques. Firstly, it can handle documents with multi-colored texts with different background shades. Secondly, the method is applicable to documents having text of widely varying sizes, usually not handled by local binarization methods. Thirdly, the method automatically computes the threshold for binarization and the logic for inverting the output from the image data and does not require any input parameter. However, the method is sensitive to complex backgrounds since it relies on the edge information to identify CCs. It also uses script-specific characteristics to filter out edge components before binarization and currently works well for Roman script only. (b) Contour-based color clustering (COCOCLUST): To overcome the above limitations, we introduce a novel unsupervised color clustering approach that operates on a ‘small’ representative set of color pixels identified using the contour information. Based on the assumption that every character is of a uniform color, we analyze each color layer individually and identify potential text regions for binarization. Experiments on several complex images having large variations in font, size, color, orientation and script illustrate the robustness of the method. 3. Multi-script and multi-oriented text extraction from scene images Scene text understanding normally involves a pre-processing step of text detection and extraction before subjecting the acquired image for character recognition task. The subsequent recognition task is performed only on the detected text regions so as to mitigate the effect of background complexity. We propose a color-based CC labeling for robust text segmentation from natural scene images. Text CCs are identified using a combination of support vector machine and neural network classifiers trained on a set of low-level features derived from the boundary, stroke and gradient information. We develop a semiautomatic annotation toolkit to generate pixel-accurate groundtruth of 100 scenic images containing text in various layout styles and multiple scripts. The overall precision, recall and f-measure obtained on our dataset are 0.8, 0.86 and 0.83, respectively. The proposed method is also compared with others in the literature using the ICDAR 2003 robust reading competition dataset, which, however, has only horizontal English text. The overall precision, recall and f-measure obtained are 0.63, 0.59 and 0.61 respectively, which is comparable to the best performing methods in the ICDAR 2005 text locating competition. A recent method proposed by Epshtein et al. [1] achieves better results but it cannot handle arbitrarily oriented text. Our method, however, works well for generic scene images having arbitrary text orientations. 4. Alignment of curved text lines Conventional OCR systems perform poorly on document images that contain multi-oriented text lines. We propose a technique that first identifies individual text lines by grouping adjacent CCs based on their proximity and regularity. For each identified text string, a B-spline curve is fitted to the centroids of the constituent characters and normal vectors are computed along the fitted curve. Each character is then individually rotated such that the corresponding normal vector is aligned with the vertical axis. The method has been tested on a data set consisting of 50 images with text laid out in various ways namely along arcs, waves, triangles and a combination of these with linearly skewed text lines. It yields 95.9% recognition accuracy on text strings, where, before alignment, state-of-the-art OCRs fail to recognize any text. The CC-based pre-processing algorithms developed are well-suited for processing camera-captured images. We demonstrate the feasibility of the algorithms on the publicly-available ICDAR 2003 robust reading competition dataset and our own database comprising camera-captured document images that contain multiple scripts and arbitrary text layouts.

Mes travaux de thèse s’inscrivent dans le cadre du projet ANR GRAPHEM1 (Graphemebased Retrieval and Analysis for PaleograpHic Expertise of Middle Age Manuscripts). Ilsprésentent une contribution méthodologique applicable à l'analyse automatique des écrituresanciennes pour assister les experts en paléographie dans le délicat travail d’étude et dedéchiffrage des écritures.L’objectif principal est de contribuer à une instrumetation du corpus des manuscritsmédiévaux détenus par l’Institut de Recherche en Histoire des Textes (IRHT – Paris) en aidantles paléographes spécialisés dans ce domaine dans leur travail de compréhension de l’évolutiondes formes de l’écriture par la mise en place de méthodes efficaces d’accès au contenu desmanuscrits reposant sur une analyse fine des formes décrites sous la formes de petits fragments(les graphèmes). Dans mes travaux de doctorats, j’ai choisi d’étudier la dynamique del’élément le plus basique de l’écriture appelé le ductus2 et qui d’après les paléographes apportebeaucoup d’informations sur le style d’écriture et l’époque d’élaboration du manuscrit.Mes contributions majeures se situent à deux niveaux : une première étape de prétraitementdes images fortement dégradées assurant une décomposition optimale des formes en graphèmescontenant l’information du ductus. Pour cette étape de décomposition des manuscrits, nousavons procédé à la mise en place d’une méthodologie complète de suivi de traits à partir del’extraction d’un squelette obtenu à partir de procédures de rehaussement de contraste et dediffusion de gradients. Le suivi complet du tracé a été obtenu à partir de l’application des règlesfondamentales d’exécution des traits d’écriture, enseignées aux copistes du Moyen Age. Il s’agitd’information de dynamique de formation des traits portant essentiellement sur des indicationsde directions privilégiées.Dans une seconde étape, nous avons cherché à caractériser ces graphèmes par desdescripteurs de formes visuelles compréhensibles à la fois par les paléographes et lesinformaticiens et garantissant une représentation la plus complète possible de l’écriture d’unpoint de vue géométrique et morphologique. A partir de cette caractérisation, nous avonsproposé une approche de clustering assurant un regroupement des graphèmes en classeshomogènes par l’utilisation d’un algorithme de classification non-supervisé basée sur lacoloration de graphe. Le résultat du clustering des graphèmes a conduit à la formation dedictionnaires de formes caractérisant de manière individuelle et discriminante chaque manuscrittraité. Nous avons également étudié la puissance discriminatoire de ces descripteurs afin d’obtenir la meilleure représentation d’un manuscrit en dictionnaire de formes. Cette étude a étéfaite en exploitant les algorithmes génétiques par leur capacité à produire de bonne sélection decaractéristiques.L’ensemble de ces contributions a été testé à partir d’une application CBIR sur trois bases demanuscrits dont deux médiévales (manuscrits de la base d’Oxford et manuscrits de l’IRHT, baseprincipale du projet), et une base comprenant de manuscrits contemporains utilisée lors de lacompétition d’identification de scripteurs d’ICDAR 2011. L’exploitation de notre méthode dedescription et de classification a été faite sur une base contemporaine afin de positionner notrecontribution par rapport aux autres travaux relevant du domaine de l’identification d’écritures etétudier son pouvoir de généralisation à d’autres types de documents. Les résultats trèsencourageants que nous avons obtenus sur les bases médiévales et la base contemporaine, ontmontré la robustesse de notre approche aux variations de formes et de styles et son caractèrerésolument généralisable à tout type de documents écrits. / My thesis work is part of the ANR GRAPHEM Project (Grapheme based Retrieval andAnalysis for Expertise paleographic Manuscripts of Middle Age). It represents a methodologicalcontribution applicable to the automatic analysis of ancient writings to assist the experts inpaleography in the delicate work of the studying and deciphering the writing.The main objective is to contribute to an instrumentation of the corpus of medievalmanuscripts held by “Institut de Recherche en Histoire de Textes” (IRHT-Paris), by helping thepaleographers specialized in this field in their work of understanding the evolution of forms inthe writing, with the establishment of effective methods to access the contents of manuscriptsbased on a fine analysis of the forms described in the form of small fragments (graphemes). Inmy PhD work, I chose to study the dynamic of the most basic element of the writing called theductus and which according to the paleographers, brings a lot of information on the style ofwriting and the era of the elaboration of the manuscript.My major contribution is situated at two levels: a first step of preprocessing of severelydegraded images to ensure an optimal decomposition of the forms into graphemes containingthe ductus information. For this decomposition step of manuscripts, we have proceeded to theestablishment of a complete methodology for the tracings of strokes by the extraction of theskeleton obtained from the contrast enhancement and the diffusion of the gradient procedures.The complete tracking of the strokes was obtained from the application of fundamentalexecution rules of the strokes taught to the scribes of the Middle Ages. It is related to thedynamic information of the formation of strokes focusing essentially on indications of theprivileged directions.In a second step, we have tried to characterize the graphemes by visual shape descriptorsunderstandable by both the computer scientists and the paleographers and thus unsuring themost complete possible representation of the wrting from a geometrical and morphological pointof view. From this characterization, we have have proposed a clustering approach insuring agrouping of graphemes into homogeneous classes by using a non-supervised classificationalgorithm based on the graph coloring. The result of the clustering of graphemes led to theformation of a codebook characterizing in an individual and discriminating way each processedmanuscript. We have also studied the discriminating power of the descriptors in order to obtaina better representation of a manuscript into a codebook. This study was done by exploiting thegenetic algorithms by their ability to produce a good feature selection.The set of the contributions was tested from a CBIR application on three databases ofmanuscripts including two medieval databases (manuscripts from the Oxford and IRHTdatabases), and database of containing contemporary manuscripts used in the writersidentification contest of ICDAR 2011. The exploitation of our description and classificationmethod was applied on a cotemporary database in order to position our contribution withrespect to other relevant works in the writrings identification domain and study itsgeneralization power to other types of manuscripts. The very encouraging results that weobtained on the medieval and contemporary databases, showed the robustness of our approachto the variations of the shapes and styles and its resolutely generalized character to all types ofhandwritten documents.

Visual place recognition is the task of finding same places in a set of database images for a given set of query images. This becomes particularly challenging for long-term applications when the environmental condition changes between or within the database and query set, e.g., from day to night. Visual place recognition in changing environments can be used if global position data like GPS is not available or very inaccurate, or for redundancy. It is required for tasks like loop closure detection in SLAM, candidate selection for global localization, or multi-robot/multi-session mapping and map merging. In contrast to pure image retrieval, visual place recognition can often build upon additional information and data for improvements in performance, runtime, or memory usage. This includes additional data-inherent knowledge about information that is contained in the image sets themselves because of the way they were recorded. Using data-inherent knowledge avoids the dependency on other sensors, which increases the generality of methods for an integration into many existing place recognition pipelines. This thesis focuses on the usage of additional data-inherent knowledge. After the discussion of basics about visual place recognition, the thesis gives a systematic overview of existing data-inherent knowledge and corresponding methods. Subsequently, the thesis concentrates on a deeper consideration and exploitation of four different types of additional data-inherent knowledge. This includes 1) sequences, i.e., the database and query set are recorded as spatio-temporal sequences so that consecutive images are also adjacent in the world, 2) knowledge of whether the environmental conditions within the database and query set are constant or continuously changing, 3) intra-database similarities between the database images, and 4) intra-query similarities between the query images. Except for sequences, all types have received only little attention in the literature so far. For the exploitation of knowledge about constant conditions within the database and query set (e.g., database: summer, query: winter), the thesis evaluates different descriptor standardization techniques. For the alternative scenario of continuous condition changes (e.g., database: sunny to rainy, query: sunny to cloudy), the thesis first investigates the qualitative and quantitative impact on the performance of image descriptors. It then proposes and evaluates four unsupervised learning methods, including our novel clustering-based descriptor standardization method K-STD and three PCA-based methods from the literature. To address the high computational effort of descriptor comparisons during place recognition, our novel method EPR for efficient place recognition is proposed. Given a query descriptor, EPR uses sequence information and intra-database similarities to identify nearly all matching descriptors in the database. For a structured combination of several sources of additional knowledge in a single graph, the thesis presents our novel graphical framework for place recognition. After the minimization of the graph's error with our proposed ICM-based optimization, the place recognition performance can be significantly improved. For an extensive experimental evaluation of all methods in this thesis and beyond, a benchmark for visual place recognition in changing environments is presented, which is composed of six datasets with thirty sequence combinations.