Dynamic programming has been widely used to solve various kinds of optimization problems.In this work, we show that two crucial problems in video-based sign language and gesture recognition systems can be attacked by dynamic programming with additional multiple observations. The first problem occurs at the higher (sentence) level. Movement epenthesis [1] (me), i.e., the necessary but meaningless movement between signs, can result in difficulties in modeling and scalability as the number of signs increases. The second problem occurs at the lower (feature) level. Ambiguity of hand detection and occlusion will propagate errors to the higher level. We construct a novel framework that can handle both of these problems based on a dynamic programming approach. The me has only be modeled explicitly in the past. Our proposed method tries to handle me in a dynamic programming framework where we model the me implicitly. We call this enhanced Level Building (eLB) algorithm. This formulation also allows the incorporation of statistical grammar models such as bigrams and trigrams. Another dynamic programming process that handles the problem of selecting among multiple hand candidates is also included in the feature level. This is different from most of the previous approaches, where a single observation is used. We also propose a grouping process that can generate multiple, overlapping hand candidates. We demonstrate our ideas on three continuous American Sign Language data sets and one hand gesture data set. The ASL data sets include one with a simple background, one with a simple background but with the signer wearing short sleeved clothes, and the last with a complex and changing background. The gesture data set contains color gloved gestures with a complex background. We achieve within 5% performance loss from the automatically chosen me score compared with the manually chosen me score. At the low level, we first over segment each frame to get a list of segments. Then we use a greedy method to group the segments based on different grouping cues. We also show that the performance loss is within 5% when we compare this method with manually selected feature vectors.

The reseach on neural networks has been done for six decades. In this period, many neural models and learning rules have been proposed. Futhermore, they were popularly and successfully applied to many applications. They successfully solved many problems that traditional algorithms could not solve efficiently . However, applying multilayer neural networks to applications, users are confronted with the problem of determining the number of hidden layers and the number of hidden neurons in each hidden layer. It is too difficult for users to determine proper neural network architectures. However, it is very significant, because neural network architectures always influence critically their performance. We may solve problems efficiently, only when we has proper neural network architectures. To overcome this difficulty, several approaches have been proposed to generate the architecture of neural networks recently. However, they still have some drawbacks. The goal of our research is to discover better approachs to automatically determine proper neural network architectures. We propose a series of approaches in this thesis. First, we propose an approach based on decision trees. It successfully determines neural network architectures and greatly decreases learning time. However, it can deal only with two-class problems and it generates bigger neural network architectures. Next, we propose an information entropy based approach to overcome the above drawbacks. It can generate easily multi-class neural networks for standard domain problems. Finally, we expand the above method for sequential domain and structured domain problems. Therefore, our approaches can be applied to many applications. Currently, we are trying to work on quantum neural networks. We are also interested in ART neural networks. They are also incremental neural models. We apply them to digital signal processing. We propose a character recognition application, a spoken word recognition application, and an image compression application. All of them have good performances.

Cette étude consiste à étudier l'apport de réseaux de neurones<br />artificiels pour améliorer le contrôle de processus industriels<br />complexes, caractérisés en particulier par leur aspect temporel.<br />Les motivations principales pour traiter des séries temporelles<br />sont la réduction du volume de données, l'indexation pour la<br />recherche de similarités, la localisation de séquences,<br />l'extraction de connaissances (data mining) ou encore la<br />prédiction.<br /><br />Le processus industriel choisi est un four à arc<br />électrique pour la production d'acier liquide au Luxembourg. Notre<br />approche est un concept de contrôle prédictif et se base sur des<br />méthodes d'apprentissage non-supervisé dans le but d'une<br />extraction de connaissances.<br /><br />Notre méthode de codage se base sur<br />des formes primitives qui composent les signaux. Ces formes,<br />composant un alphabet de codage, sont extraites par une méthode<br />non-supervisée, les cartes auto-organisatrices de Kohonen (SOM).<br />Une méthode de validation des alphabets de codage accompagne<br />l'approche.<br /><br />Un sujet important abordé durant ces recherches est<br />la similarité de séries temporelles. La méthode proposée est<br />non-supervisée et intègre la capacité de traiter des séquences de<br />tailles variées.

Dynamic speech properties such as time warping, silence removal and background noise interference are the most challenging issues in continuous speech signal matching. Among all of them, the time warped speech signal matching is of great interest and has been a tough challenge for the researchers. An adaptive framing based continuous speech tracking and similarity measurement approach is introduced in this work following a comprehensive research conducted in the diverse areas of speech processing. A dynamic state model is introduced based on system of linear motion equations which models the input (test) speech signal frame as a unidirectional moving object along the template speech signal. The most similar corresponding frame position in the template speech is estimated which is fused with a feature based similarity observation and the noise variances using a Kalman filter. The Kalman filter provides the final estimated frame position in the template speech at current time which is further used for prediction of a new frame size for the next step. In addition, a keyword spotting approach is proposed by introducing wavelet decomposition based dynamic noise filter and combination of beliefs. The Dempster’s theory of belief combination is deployed for the first time in relation to keyword spotting task. Performances for both; speech tracking and keyword spotting approaches are evaluated using the statistical metrics and gold standards for the binary classification. Experimental results proved the superiority of the proposed approaches over the existing methods.

Process operators often have process faults and alarms due to recurring failures on process equipment. It is also the case that some processes do not have enough input information or process models to use conventional modelling or machine learning techniques for early fault detection. A proof of concept for online streaming prediction software based on matching process behaviour to historical motifs has been developed, making use of the Basic Local Alignment Search Tool (BLAST) used in the Bioinformatics field. Execution times of as low as 1 second have been recorded, demonstrating that online matching is feasible. Three techniques have been tested and compared in terms of their computational effciency, robustness and selectivity, with results shown in Table 1: • Symbolic Aggregate Approximation combined with PSI-BLAST • Naive Triangular Representation with PSI-BLAST • Dynamic Time Warping Table 1: Properties of different motif-matching methods Property SAX-PSIBLAST TER-PSIBLAST DTW Noise tolerance (Selectivity) Acceptable Inconclusive Good Vertical Shift tolerance None Perfect Poor Matching speed Acceptable Acceptable Fast Match speed scaling O < O(mn) O < O(mn) O(mn) Dimensionality Reduction Tolerance Good Inconclusive Acceptable It is recommended that a method using a weighted confidence measure for each technique be investigated for the purpose of online process event handling and operator alerts. Keywords: SAX, BLAST, motif-matching, Dynamic Time Warping / Dissertation (MEng)--University of Pretoria, 2012. / Chemical Engineering / unrestricted

Dynamic speech properties such as time warping, silence removal and background noise interference are the most challenging issues in continuous speech signal matching. Among all of them, the time warped speech signal matching is of great interest and has been a tough challenge for the researchers. An adaptive framing based continuous speech tracking and similarity measurement approach is introduced in this work following a comprehensive research conducted in the diverse areas of speech processing. A dynamic state model is introduced based on system of linear motion equations which models the input (test) speech signal frame as a unidirectional moving object along the template speech signal. The most similar corresponding frame position in the template speech is estimated which is fused with a feature based similarity observation and the noise variances using a Kalman filter. The Kalman filter provides the final estimated frame position in the template speech at current time which is further used for prediction of a new frame size for the next step. In addition, a keyword spotting approach is proposed by introducing wavelet decomposition based dynamic noise filter and combination of beliefs. The Dempster’s theory of belief combination is deployed for the first time in relation to keyword spotting task. Performances for both; speech tracking and keyword spotting approaches are evaluated using the statistical metrics and gold standards for the binary classification. Experimental results proved the superiority of the proposed approaches over the existing methods. / The appendices files are not available online.

For a long time, handwriting analysis, such as handwriting recognition and signature verification, has been an active research area. There are two categories of handwriting, online and offline. Online handwriting is captured in real-time on a digital device such as a tablet screen with a stylus pen. In contrast, the handwritten text scanned or captured by a camera from a physical medium such as paper is referred to as offline handwriting. For offline handwriting, the input is limited to handwritten images, making handwriting analysis much more difficult. In our work, we proposed a Stroke Trajectory Recover (STR) for offline and unconstrained handwritten documents. For this purpose, we introduce large-scale word-level annotations for the English handwriting sampled from the IAM-online dataset. The current STR architectures for English handwriting use lines of text or characters of the alphabet as input. However, a word-level STR method estimates loss for each word rather than averaging DTW loss over the entire line of text. Furthermore, to avoid the stray points/artifacts in predicted stroke points, we employ a marginal Chamfer distance that penalizes large, easily noticeable deviations and artifacts. For word detection, we propose the fusion of character region scores with bounding box estimation. Since the character level annotations are not available for handwritten text, we estimate the character region scores in a weakly supervised manner. Character region scores are estimated autonomously from the word’s bounding box estimation to learn the character level information in handwriting. We propose to fuse the character region scores and images to detect words in camera-captured handwriting images. We also propose an automated evaluation to check the quality of the predicted stroke trajectory. The existing handwriting datasets have limited availability of stroke coordinates information. Hence, although the proposed system can be applied to handwriting datasets without stroke coordinates information, it is impossible to evaluate the quality of its predicted strokes using the existing methods. Therefore, in our work, we propose two measures for evaluating the quality of recovered stroke trajectories when ground truth stroke information is not given. First, we formulated an automated evaluation measure based on image matching by finding the difference between original and rendered images. We also evaluated the preservation of readability of words for original and rendered images with a transformer-based word recognition network. Since our proposed STR system works with words, we demonstrate that our method is scalable to unconstrained handwritten documents, i.e., full-page text. Finally, we present a probabilistic diffusion model conditioned on handwriting style template for generating writing strokes. In our work, we propose to learn the localized patches for handwriting style features from multiscale attention network. The multiscale attention network captures fine details about local character style and global handwriting style. Moreover, we train our diffusion model with the Dynamic Time Warping (DTW) loss function, along with the diffusion loss, which eliminates the need to train any auxiliary networks for text or writer style recognition and adversarial networks. / Computer and Information Science

Submitted by PPG Ci?ncia da Computa??o (ppgcc@pucrs.br) on 2018-06-19T12:25:43Z No. of bitstreams: 1 GRAZIELE_MARQUES_MAZUCO_DOS_SANTOS_DIS.pdf: 3857481 bytes, checksum: 9c3c88f01e8e5d920ba3bc8989d2cfbf (MD5) / Approved for entry into archive by Sheila Dias (sheila.dias@pucrs.br) on 2018-06-27T13:05:50Z (GMT) No. of bitstreams: 1 GRAZIELE_MARQUES_MAZUCO_DOS_SANTOS_DIS.pdf: 3857481 bytes, checksum: 9c3c88f01e8e5d920ba3bc8989d2cfbf (MD5) / Made available in DSpace on 2018-06-27T13:21:15Z (GMT). No. of bitstreams: 1 GRAZIELE_MARQUES_MAZUCO_DOS_SANTOS_DIS.pdf: 3857481 bytes, checksum: 9c3c88f01e8e5d920ba3bc8989d2cfbf (MD5) Previous issue date: 2018-04-26 / Demand forecasting is one of the most essential components of supply chain management. Forecasts are used both for long-term and for short-term. Long-term forecasts are important because it is difficult in terms of production to face the demand deviation in a short time, so the anticipation of prediction helps to increase the responsiveness of the supply chain. Short term forecasts are important for the demand monitoring aiming to keep healthy inventory levels. In the fashion industry, the high change of products, the short life cycle and the lack of historical data makes difficult accurate predictions. To deal with this problem, the literature presents three approaches: statistical, artificial intelligence and hybrid that combines statistical and artificial intelligence. This research presents a two-phased method: (1) long-term prediction, identifies the different life cycles in the products, allowing the identification of sales prototypes for each cluster and (2) short-term prediction, classifies new products in the clusters labeled in the long-term phase and adjusts the sales curve considering optimistic and pessimist factors. As a differential, the method is based in dynamic time warping, distance measure for time series. The method is tested in a real dataset with real data from fashion retailers that demonstrates the quality of the contribution. / A previs?o de vendas no varejo da moda ? um problema complexo e um dos componentes essenciais da cadeia de suprimento, sendo utilizada tanto para previs?o de longo prazo quanto para a previs?o de curto prazo. A previs?o de longo prazo ? importante pois ? dif?cil, em termos de produ??o, enfrentar o desvio da demanda em um curto espa?o de tempo, ent?o a previs?o antecipada permite aumentar a capacidade de resposta da cadeia de suprimento. A previs?o de curto prazo ? importante para o acompanhamento da demanda, visando a adequa??o do n?vel de estoque. No varejo da moda a alta rotatividade, o curto ciclo de vida dos produtos e a consequente aus?ncia de dados hist?ricos dificulta a gera??o de previs?es precisas. Para lidar com esse problema, h? na literatura tr?s principais abordagens: estat?stica, baseada em intelig?ncia artificial e h?brida, que combina estat?stica e intelig?ncia artificial. Esta pesquisa prop?e um m?todo de previs?o de vendas em duas etapas: (1) previs?o de longo prazo, que pretende detectar diferentes grupos de produtos com ciclos de vida semelhantes, permitindo assim a identifica??o do comportamento m?dio de cada um dos grupos e (2) previs?o de curto prazo que busca associar os produtos novos nos grupos identificados na etapa de longo prazo e ajustar a curva de vendas levando em considera??o fatores conservadores, otimistas ou pessimistas. Al?m disso, nesta etapa ? poss?vel realizar a previs?o de reposi??o de itens. Como diferencial, o m?todo proposto utiliza a medida de dist?ncia Dynamic Time Warping, identificada na literatura como adequada para lidar com s?ries temporais. O m?todo ? testado utilizando dois conjuntos de dados reais de varejistas da moda, foram realizados dois experimentos, que demonstram a qualidade da contribui??o.

Notre travail porte sur l'extraction de connaissances sur des langages graphiques dont les symboles sont a priori inconnus. Nous formons l'hypothèse que l'observation d'une grande quantité de documents doit permettre de découvrir les symboles composant l'alphabet du langage considéré. La difficulté du problème réside dans la nature bidimensionnelle et manuscrite des langages graphiques étudiés. Nous nous plaçons dans le cadre de tracés en-ligne produit par des interfaces de saisie de type écrans tactiles, tableaux interactifs ou stylos électroniques. Le signal disponible est alors une trajectoire échantillonnée produisant une séquence de traits, eux-mêmes composés d'une séquence de points. Un symbole, élément de base de l'alphabet du langage, est donc composé d'un ensemble de traits possédant des propriétés structurelles et relationnelles spécifiques. L'extraction des symboles est réalisée par la découverte de sous-graphes répétitifs dans un graphe global modélisant les traits (noeuds) et leur relations spatiales (arcs) de l'ensemble des documents. Le principe de description de longueur minimum (MDL : Minimum Description Length) est mis en oeuvre pour choisir les meilleurs représentants du lexique des symboles. Ces travaux ont été validés sur deux bases expérimentales. La première est une base d'expressions mathématiques simples, la seconde représente des graphiques de type organigramme. Sur ces bases, nous pouvons évaluer la qualité des symboles extraits et comparer à la vérité terrain. Enfin, nous nous sommes intéressés à la réduction de la tâche d'annotation d'une base en considérant à la fois les problématiques de segmentation et d'étiquetage des différents traits.

The rapid development of corpus-based speech systems such as concatenative synthesis systems for under-resourced languages requires an efficient, consistent and accurate solution with regard to phonetic speech segmentation. Manual development of phonetically annotated corpora is a time consuming and expensive process which suffers from challenges regarding consistency and reproducibility, while automation of this process has only been satisfactorily demonstrated on large corpora of a select few languages by employing techniques requiring extensive and specialised resources. In this work we considered the problem of phonetic segmentation in the context of developing small prototypical speech synthesis corpora for new under-resourced languages. This was done through an empirical evaluation of existing segmentation techniques on typical speech corpora in three South African languages. In this process, the performance of these techniques were characterised under different data conditions and the efficient application of these techniques were investigated in order to improve the accuracy of resulting phonetic alignments. We found that the application of baseline speaker-specific Hidden Markov Models results in relatively robust and accurate alignments even under extremely limited data conditions and demonstrated how such models can be developed and applied efficiently in this context. The result is segmentation of sufficient quality for synthesis applications, with the quality of alignments comparable to manual segmentation efforts in this context. Finally, possibilities for further automated refinement of phonetic alignments were investigated and an efficient corpus development strategy was proposed with suggestions for further work in this direction. / Thesis (M.Ing. (Computer Engineering))--North-West University, Potchefstroom Campus, 2009.