Often there is a need to establish and understand the structure of multidimensional data: their clusters, outliers, similarity and dissimilarity. One of solution ways is a dimensionality reduction and visualization of the data. If a huge datasets is analyzed, it is purposeful to reduce the number of the data items before visualization. The area of research is reduction of the number of the data analyzed and mapping the data in a plane. In the dissertation, vector quantization methods, based on artificial neural networks, and visualization methods, based on a dimensionality reduction, have been investigated. The consecutive and integrated combinations of neural gas and multidimensional scaling have been proposed here as an alternative to combinations of self-organizing maps and multidimensional scaling. The visualization quality is estimated by König’s topology preservation measure, Spearman’s rho and MDS error. The measures allow us to evaluate the similarity preservation quantitatively after a transformation of multidimensional data into a lower dimension space. The ways of selecting the initial values of two-dimensional vectors in the consecutive combination and the first training block of the integrated combination have been proposed and the ways of assigning the initial values of two-dimensional vectors in all the training blocks, except the first one, of the integrated combination have been developed. The dependence of the quantization error on the values of training... [to full text] / Dažnai iškyla būtinybė nustatyti ir giliau pažinti daugiamačių duomenų struktūrą: susidariusius klasterius, itin išsiskiriančius objektus, objektų tarpusavio panašumą ir skirtingumą. Vienas iš sprendimų būdų – duomenų dimensijos mažinimas ir jų vizualizavimas. Kai analizuojamos didelės duomenų aibės, tikslinga prieš vizualizavimą sumažinti ne tik dimensiją, bet ir duomenų skaičių. Šio darbo tyrimų sritis yra daugiamačių duomenų skaičiaus mažinimas ir duomenų atvaizdavimas plokštumoje. Disertacijoje nagrinėjami dirbtiniais neuroniniais tinklais grindžiami vektorių kvantavimo ir dimensijos mažinimu pagrįsti vizualizavimo metodai. Kaip alternatyva saviorganizuojančių neuroninių tinklų ir daugiamačių skalių junginiams, darbe pasiūlyti nuoseklus neuroninių dujų ir daugiamačių skalių junginys bei integruotas, atsižvelgiantis į neuroninių dujų metodo mokymosi eigą ir leidžiantis gauti tikslesnę daugiamačių vektorių projekciją plokštumoje. Junginiais gautų vaizdų kokybės vertinimui pasirinkti Konigo matas, Spirmano koeficientas bei MDS paklaida. Šie matai leidžia kiekybiškai įvertinti panašumų išlaikymą po daugiamačių duomenų transformavimo į mažesnės dimensijos erdvę. Taip pat pasiūlyti dvimačių vektorių pradinių koordinačių parinkimo būdai nuosekliame junginyje ir integruoto junginio pirmame mokymo bloke bei koordinačių reikšmių priskyrimo būdai integruoto junginio kituose mokymo blokuose. Eksperimentiškai nustatyta kvantavimo paklaidos priklausomybė nuo neuroninių dujų tinklo... [toliau žr. visą tekstą]

Dažnai iškyla būtinybė nustatyti ir giliau pažinti daugiamačių duomenų struktūrą: susidariusius klasterius, itin išsiskiriančius objektus, objektų tarpusavio panašumą ir skirtingumą. Vienas iš sprendimų būdų – duomenų dimensijos mažinimas ir jų vizualizavimas. Kai analizuojamos didelės duomenų aibės, tikslinga prieš vizualizavimą sumažinti ne tik dimensiją, bet ir duomenų skaičių. Šio darbo tyrimų sritis yra daugiamačių duomenų skaičiaus mažinimas ir duomenų atvaizdavimas plokštumoje. Disertacijoje nagrinėjami dirbtiniais neuroniniais tinklais grindžiami vektorių kvantavimo ir dimensijos mažinimu pagrįsti vizualizavimo metodai. Kaip alternatyva saviorganizuojančių neuroninių tinklų ir daugiamačių skalių junginiams, darbe pasiūlyti nuoseklus neuroninių dujų ir daugiamačių skalių junginys bei integruotas, atsižvelgiantis į neuroninių dujų metodo mokymosi eigą ir leidžiantis gauti tikslesnę daugiamačių vektorių projekciją plokštumoje. Junginiais gautų vaizdų kokybės vertinimui pasirinkti Konigo matas, Spirmano koeficientas bei MDS paklaida. Šie matai leidžia kiekybiškai įvertinti panašumų išlaikymą po daugiamačių duomenų transformavimo į mažesnės dimensijos erdvę. Taip pat pasiūlyti dvimačių vektorių pradinių koordinačių parinkimo būdai nuosekliame junginyje ir integruoto junginio pirmame mokymo bloke bei koordinačių reikšmių priskyrimo būdai integruoto junginio kituose mokymo blokuose. Eksperimentiškai nustatyta kvantavimo paklaidos priklausomybė nuo neuroninių dujų tinklo... [toliau žr. visą tekstą] / Often there is a need to establish and understand the structure of multidimensional data: their clusters, outliers, similarity and dissimilarity. One of solution ways is a dimensionality reduction and visualization of the data. If a huge datasets is analyzed, it is purposeful to reduce the number of the data items before visualization. The area of research is reduction of the number of the data analyzed and mapping the data in a plane. In the dissertation, vector quantization methods, based on artificial neural networks, and visualization methods, based on a dimensionality reduction, have been investigated. The consecutive and integrated combinations of neural gas and multidimensional scaling have been proposed here as an alternative to combinations of self-organizing maps and multidimensional scaling. The visualization quality is estimated by König’s topology preservation measure, Spearman’s rho and MDS error. The measures allow us to evaluate the similarity preservation quantitatively after a transformation of multidimensional data into a lower dimension space. The ways of selecting the initial values of two-dimensional vectors in the consecutive combination and the first training block of the integrated combination have been proposed and the ways of assigning the initial values of two-dimensional vectors in all the training blocks, except the first one, of the integrated combination have been developed. The dependence of the quantization error on the values of training... [to full text]

Residual vector quantization (RVQ) is a 1-nearest neighbor (1-NN) type of technique. RVQ is a multi-stage implementation of regular vector quantization. An input is successively quantized to the nearest codevector in each stage codebook. In classification, nearest neighbor techniques are very attractive since these techniques very accurately model the ideal Bayes class boundaries. However, nearest neighbor classification techniques require a large size of representative dataset. Since in such techniques a test input is assigned a class membership after an exhaustive search the entire training set, a reasonably large training set can make the implementation cost of the nearest neighbor classifier unfeasibly costly. Although, the k-d tree structure offers a far more efficient implementation of 1-NN search, however, the cost of storing the data points can become prohibitive, especially in higher dimensionality. RVQ also offers a nice solution to a cost-effective implementation of 1-NN-based classification. Because of the direct-sum structure of the RVQ codebook, the memory and computational of cost 1-NN-based system is greatly reduced. Although, as compared to an equivalent 1-NN system, the multi-stage implementation of the RVQ codebook compromises the accuracy of the class boundaries, yet the classification error has been empirically shown to be within 3% to 4% of the performance of an equivalent 1-NN-based classifier.

The task of automatically recognizing targets in IR imagery has a history of approximately 25 years of research and development. ATR is an application of pattern recognition and scene analysis in the field of defense industry and it is still one of the challenging problems. This thesis may be viewed as an exploratory study of ATR problem with encouraging recognition algorithms implemented in the area. The examined algorithms are among the solutions to the ATR problem, which are reported to have good performance in the literature. Throughout the study, PCA, subspace LDA, ICA, nearest mean classifier, K nearest neighbors classifier, nearest neighbor classifier, LVQ classifier are implemented and their performances are compared in the aspect of recognition rate. According to the simulation results, the system, which uses the ICA as the feature extractor and LVQ as the classifier, has the best performing results. The good performance of this system is due to the higher order statistics of the data and the success of LVQ in modifying the decision boundaries.

nÃo hà / Nesta tese aborda-se o problema de identificaÃÃo de sistemas dinÃmicos sobre a Ãtica dos modelos locais, em que o espaÃo de entrada à particionado em regiÃes de operaÃÃo menores sobre as quais sÃo construÃdos modelos de menor complexidade (em geral, lineares). Este tipo de modelo à uma alternativa aos chamados modelos globais em que a dinÃmica do sistema à identificada usando-se uma Ãnica estrutura (em geral, nÃo-linear) que cobre todo o espaÃo de entrada. Assim, o tema alvo desta tese à o projeto de modelos lineares locais cujo espaÃo de entrada à particionado por meio do uso de algoritmos de quantizaÃÃo vetorial, principalmente aqueles baseados em redes neurais competitivas. Para este fim, sÃo propostos trÃs novos modelos lineares locais baseados na rede SOM (self-organizing map), que sÃo avaliados na tarefa de identificaÃÃo do modelo inverso de quatro sistemas dinÃmicos comumente usados na literatura em benchmarks de desempenhos. Os modelos propostos sÃo tambÃm comparados com modelos globais baseados nas redes MLP (multilayer perceptron) e ELM (extreme learning machines), bem como com outros modelos lineares locais, tais como o modelo fuzzy Takagi-Sugeno e o modelo neural LLM (local linear mapping). Um amplo estudo à realizado visando comparar os desempenhos de todos os modelos supracitados segundo trÃs critÃrios de avaliaÃÃo, a saber: (i) erro mÃdio quadrÃtico normalizado, (ii) anÃlise dos resÃduos, e (iii) teste estatÃstico de Kolmogorov-Smirnov. De particular interesse para esta tese, à a avaliaÃÃo da robustez dos modelos locais propostos com relaÃÃo ao algoritmo de quantizaÃÃo vetorial usado no treinamento do modelo. Os resultados obtidos indicam que os desempenhos dos modelos locais propostos sÃo superiores aos dos modelos globais baseados na rede MLP e equivalentes aos modelos globais baseados na rede ELM. / In this thesis the problem of nonlinear system identification is approached from the viewpoint of local models. The input space is partitioned into smaller operational regions with lower complexity models (usually linear) built for each one. This type of model is an alternative to global models, for which the system dynamics is identified using a single structure (usually nonlinear ones) that covers the whole input space. The aim of this thesis is to design of local linear models whose input space is partitioned by means of vector quantization algorithms, special those based on competitive learning neural networks. For this purpose, three novel local linear modeling methods based on the SOM (self-organizing map) are introduced and evaluated on the identification of the inverse model of four dynamical systems commonly used in the literature for performance benchmarking. The proposed models are also compared with global models based on the MLP (multilayer perceptron) and ELM (extreme learning machines), as well as with alternative local linear models, such as the Takagi-Sugeno fuzzy model and the LLM(local linear mapping) neural model. A comprehensive study is carried out to compare the performances of all the aforementioned models according to three evaluation criteria, namely: (i) normalized mean squared error, (ii) residual analysis, and (iii) Kolmogorov-Smirnov test. Of particular interest to this thesis is the evaluation of the robustness of the proposed local models with respect to the vector quantization algorithm used to train the model. The obtained results indicates that the performance of the proposed local models are superior to those achieved by the MLP-based global models and equivalent to those achieved by ELM-based global models.

[pt] Nesta dissertação, propomos uma metodologia para seleção de dados em modelos de Aprendizado por Quantização Vetorial, referenciado amplamente na literatura pela sigla em inglês LVQ. Treinar um modelo (ajuste dentro-daamostra) com um subconjunto selecionado a partir do conjunto de dados disponíveis para o aprendizado pode trazer grandes benefícios no resultado de generalização (fora-da-amostra). Neste sentido, é muito importante realizar uma busca para selecionar dados que, além de serem representativos de suas distribuições originais, não sejam ruído (no sentido definido ao longo desta dissertação). O método proposto procura encontrar os pontos relevantes do conjunto de entrada, tendo como base a correlação do erro de cada ponto com o erro do restante da distribuição. Procura-se, em geral, eliminar considerável parte do ruído mantendo os pontos que são relevantes para o ajuste do modelo (aprendizado). Assim, especificamente em LVQ, a atualização dos protótipos durante o aprendizado é realizada com um subconjunto do conjunto de treinamento originalmente disponível. Experimentos numéricos foram realizados com dados simulados e reais, e os resultados obtidos foram muito interessantes, mostrando claramente a potencialidade do método proposto. / [en] In this dissertation, we consider a methodology for selection of data in models of Learning Vector Quantization (LVQ). The generalization can be improved by using a subgroup selected from the available data set. We search the original distribution to select relevant data that aren't noise. The search aims at relevant points in the training set based on the correlation between the error of each point and the average of error of the remaining data. In general, it is desired to eliminate a considerable part of the noise, keeping the points that are relevant for the learning model. Thus, specifically in LVQ, the method updates the prototypes with a subgroup of the originally available training set. Numerical experiments have been done with simulated and real data. The results were very interesting and clearly indicated the potential of the method.

No description available.

No description available.

The programme for the segmentation of a speech into fonems was created as a part of the master´s thesis. This programme was made in the programme Matlab and consists of several scripts. The programme serves for automatic segmentation. Speech segmentation is the process of identifying the boundaries between phonemes in spoken natural languages. Automatic segmentation is based on vector quantization. In the first step of algorithm, feature extraction is realized. Then speech segments are assigned to calculated centroids. Position where centroid is changed is marked as a boundary of phoneme. The audiorecords were elaborated by the programme and a operation of the automatic segmentation was analysed. A detailed manual was created to the programme too. Individual used methods of the elaboration of a speech were in the master´s thesis briefly descripted, its implementations in the programme and reasons of set of its parameters.

Depuis plus de vingt ans, l'accès à la technologie est devenu très facile étant donné son omniprésence dans le quotidien de chacun et son faible coût. Cet accès aux technologies du numérique permet à toute personne équipée d'un ordinateur ou d'un smartphone de visualiser et de modifier des contenus digitaux. Avec les progrès en matière de stockage en ligne, la quantité de contenus digitaux tels que le son, l'image ou la vidéo sur internet a explosé et continue d'augmenter.Savoir identifier la source d'une image et certifier si celle-ci a été modifiée ou non sont des informations nécessaires pour authentifier une image et ainsi protéger la propriété intellectuelle et les droits d’auteur par exemple. Une des approches pour résoudre ces problèmes est le tatouage numérique. Il consiste à insérer une marque dans une image qui permettra de l'authentifier.Dans cette thèse, nous étudions premièrement le tatouage numérique dans le but de proposer des méthodes plus robustes aux modifications d'image grâce aux codes correcteurs. En étudiant la structure des erreurs produites par la modification d’une image marquée, un code correcteur sera plus efficace qu’un autre. Nous proposons aussi d’intégrer de nouveaux codes correcteurs appelés codes en métrique rang pour le tatouage.Ensuite, nous proposons d’améliorer l'invisibilité des méthodes de tatouage pour les images couleur. A l’insertion d’une marque, les dégradations de l’image sont perçues différemment par le système visuel humain en fonction de la couleur. Nous proposons un modèle biologique de la perception des couleurs qui nous permet de minimiser les distorsions psychovisuelles de l’image à l’insertion.Toutes ces techniques sont testées sur des images naturelles dans un contexte d’insertion d’information. / For more than twenty years, technology has become more and more easy to access. It is omnipresent in everyday life and is low cost. It allows anyone using a computer or a smartphone to visualize and modify digital contents. Also, with the impressive progress of online massive data storage (cloud), the quantity of digital contents has soared and continues to increase. To ensure the protection of intellectual property and copyright, knowing if an image has been modified or not is an important information in order to authenticate it. One approach to protect digital contents is digital watermarking. It consists in modifying an image to embed an invisible mark which can authenticate the image. In this doctorate thesis, we first study how to improve the robustness of digital image watermarking against image processings thanks to error correcting codes. By studying the error structure produced by the image processing applied on a watermarked image, we can find an optimal choice of error correcting code for the best correction performances. Also, we propose to integrate a new type of error correcting codes called rank metric codes for watermarking applications. Then, we propose to improve the invisibility of color image watermarking methods. At the embedding step, a host image suffers some distortions which are perceived differently in function of the color by the human visual system. We propose a biological model of color perception which allows one to minimize psychovisual distortions applied on the image to protect.