Learning Embeddings for Indexing, Retrieval, and Classification, with Applications to Object and Shape Recognition in Image Databases

Nearest neighbor retrieval is the task of identifying, given a database of objects and a query object, the objects in the database that are the most similar to the query. Retrieving nearest neighbors is a necessary component of many practical applications, in fields as diverse as computer vision, pattern recognition, multimedia databases, bioinformatics, and computer networks. At the same time, finding nearest neighbors accurately and efficiently can be challenging, especially when the database contains a large number of objects, and when the underlying distance measure is computationally expensive. This thesis proposes new methods for improving the efficiency and accuracy of nearest neighbor retrieval and classification in spaces with computationally expensive distance measures. The proposed methods are domain-independent, and can be applied in arbitrary spaces, including non-Euclidean and non-metric spaces. In this thesis particular emphasis is given to computer vision applications related to object and shape recognition, where expensive non-Euclidean distance measures are often needed to achieve high accuracy. The first contribution of this thesis is the BoostMap algorithm for embedding arbitrary spaces into a vector space with a computationally efficient distance measure. Using this approach, an approximate set of nearest neighbors can be retrieved efficiently - often orders of magnitude faster than retrieval using the exact distance measure in the original space. The BoostMap algorithm has two key distinguishing features with respect to existing embedding methods. First, embedding construction explicitly maximizes the amount of nearest neighbor information preserved by the embedding. Second, embedding construction is treated as a machine learning problem, in contrast to existing methods that are based on geometric considerations. The second contribution is a method for constructing query-sensitive distance measures for the purposes of nearest neighbor retrieval and classification. In high-dimensional spaces, query-sensitive distance measures allow for automatic selection of the dimensions that are the most informative for each specific query object. It is shown theoretically and experimentally that query-sensitivity increases the modeling power of embeddings, allowing embeddings to capture a larger amount of the nearest neighbor structure of the original space. The third contribution is a method for speeding up nearest neighbor classification by combining multiple embedding-based nearest neighbor classifiers in a cascade. In a cascade, computationally efficient classifiers are used to quickly classify easy cases, and classifiers that are more computationally expensive and also more accurate are only applied to objects that are harder to classify. An interesting property of the proposed cascade method is that, under certain conditions, classification time actually decreases as the size of the database increases, a behavior that is in stark contrast to the behavior of typical nearest neighbor classification systems. The proposed methods are evaluated experimentally in several different applications: hand shape recognition, off-line character recognition, online character recognition, and efficient retrieval of time series. In all datasets, the proposed methods lead to significant improvements in accuracy and efficiency compared to existing state-of-the-art methods. In some datasets, the general-purpose methods introduced in this thesis even outperform domain-specific methods that have been custom-designed for such datasets.

An automated machine vision system for the histological grading of cervical intraepithelial neoplasia (CIN)

The histological grading of cervical intraepithelial neoplasia (CIN) remains subjective, resulting in inter- and intra-observer variation and poor reproducibility in the grading of cervical lesions. This study has attempted to develop an objective grading system using automated machine vision. The architectural features of cervical squamous epithelium are quantitatively analysed using a combination of computerized digital image processing and Delaunay triangulation analysis; 230 images digitally captured from cases previously classified by a gynaecological pathologist included normal cervical squamous epithelium (n = 30), koilocytosis (n = 46), CIN 1 (n = 52), CIN 2 (n = 56), and CIN 3 (n=46). Intra- and inter-observer variation had kappa values of 0.502 and 0.415, respectively. A machine vision system was developed in KS400 macro programming language to segment and mark the centres of all nuclei within the epithelium. By object-oriented analysis of image components, the positional information of nuclei was used to construct a Delaunay triangulation mesh. Each mesh was analysed to compute triangle dimensions including the mean triangle area, the mean triangle edge length, and the number of triangles per unit area, giving an individual quantitative profile of measurements for each case. Discriminant analysis of the geometric data revealed the significant discriminatory variables from which a classification score was derived. The scoring system distinguished between normal and CIN 3 in 98.7% of cases and between koilocytosis and CIN 1 in 76.5% of cases, but only 62.3% of the CIN cases were classified into the correct group, with the CIN 2 group showing the highest rate of misclassification. Graphical plots of triangulation data demonstrated the continuum of morphological change from normal squamous epithelium to the highest grade of CIN, with overlapping of the groups originally defined by the pathologists. This study shows that automated location of nuclei in cervical biopsies using computerized image analysis is possible. Analysis of positional information enables quantitative evaluation of architectural features in CIN using Delaunay triangulation meshes, which is effective in the objective classification of CIN. This demonstrates the future potential of automated machine vision systems in diagnostic histopathology. Copyright (C) 2000 John Wiley and Sons, Ltd.

A Cyclotomic Lattice Based Quasi-Orthogonal STBC for Eight Transmit Antennas

In this letter, we propose a lattice-based full diversity design for rate-one quasi-orthogonal space time block codes (QSTBC) to obtain an improved diversity product for eight transmit antennas where the information bits are mapped into 4-D lattice points instead of the common modulation constellations. Particularly, the diversity product of the proposed code is directly determined by the minimum Euclidean distance of the used lattice and can be improved by using the lattice packing. We show analytically and by using simulation results that the proposed code achieves a larger diversity product than the rate-one QSTBCs reported previously.

A bit-level systolic architecture for implementing a VQ tree search

A bit-level systolic array system for performing a binary tree Vector Quantization codebook search is described. This consists of a linear chain of regular VLSI building blocks and exhibits data rates suitable for a wide range of real-time applications. A technique is described which reduces the computation required at each node in the binary tree to that of a single inner product operation. This method applies to all the common distortion measures (including the Euclidean distance, the Weighted Euclidean distance and the Itakura-Saito distortion measure) and significantly reduces the hardware required to implement the tree search system. © 1990 Kluwer Academic Publishers.

Clustering aplicado à Bolsa de Valores de Lisboa

O objetivo desta dissertação foi estudar um conjunto de empresas cotadas na bolsa de valores de Lisboa, para identificar aquelas que têm um comportamento semelhante ao longo do tempo. Para isso utilizamos algoritmos de Clustering tais como K-Means, PAM, Modelos hierárquicos, Funny e C-Means tanto com a distância euclidiana como com a distância de Manhattan. Para selecionar o melhor número de clusters identificado por cada um dos algoritmos testados, recorremos a alguns índices de avaliação/validação de clusters como o Davies Bouldin e Calinski-Harabasz entre outros.

Conformational analysis of antibody binding site using EDMA methods /

Euclidean distance matrix analysis (EDMA) methods are used to distinguish whether or not significant difference exists between conformational samples of antibody complementarity determining region (CDR) loops, isolated LI loop and LI in three-loop assembly (LI, L3 and H3) obtained from Monte Carlo simulation. After the significant difference is detected, the specific inter-Ca distance which contributes to the difference is identified using EDMA.The estimated and improved mean forms of the conformational samples of isolated LI loop and LI loop in three-loop assembly, CDR loops of antibody binding site, are described using EDMA and distance geometry (DGEOM). To the best of our knowledge, it is the first time the EDMA methods are used to analyze conformational samples of molecules obtained from Monte Carlo simulations. Therefore, validations of the EDMA methods using both positive control and negative control tests for the conformational samples of isolated LI loop and LI in three-loop assembly must be done. The EDMA-I bootstrap null hypothesis tests showed false positive results for the comparison of six samples of the isolated LI loop and true positive results for comparison of conformational samples of isolated LI loop and LI in three-loop assembly. The bootstrap confidence interval tests revealed true negative results for comparisons of six samples of the isolated LI loop, and false negative results for the conformational comparisons between isolated LI loop and LI in three-loop assembly. Different conformational sample sizes are further explored by combining the samples of isolated LI loop to increase the sample size, or by clustering the sample using self-organizing map (SOM) to narrow the conformational distribution of the samples being comparedmolecular conformations. However, there is no improvement made for both bootstrap null hypothesis and confidence interval tests. These results show that more work is required before EDMA methods can be used reliably as a method for comparison of samples obtained by Monte Carlo simulations.

La route qui mène au crime : déterminants de la mobilité des infracteurs de Gatineau en 2006

Cette étude porte sur la distance parcourue pour commettre un crime à Gatineau en 2006. Peu d’études canadiennes récentes ont porté sur le sujet. De plus, il existe un vide de connaissances sur la mobilité des délinquants dans les petites villes et les banlieues. La présente recherche vise à comparer trois mesures de distance différentes, à vérifier si la distance parcourue varie en fonction du type de crime et à voir si les variables de temps (jour de la semaine, moment de la journée et saison) de même que certaines caractéristiques des suspects (âge, sexe et lieu de résidence) ont un impact sur la distance parcourue. Pour chaque crime, l’adresse du suspect et le lieu du crime ont été géocodées pour ensuite calculer la distance entre les deux points. Il ressort de l’analyse de la forme des courbes de distances que seules les agressions sexuelles présentent une zone tampon. Les résultats des analyses statistiques indiquent que les jeunes sont plus mobiles que les suspects plus âgés et que les hommes parcourent une distance plus élevée que les femmes. Étonnement, la distance parcourue ne diffère pas significativement selon la saison et le moment de la journée. Enfin, comparativement aux autres criminels, les délinquants qui ont commis un vol qualifié sont ceux qui ont parcouru les plus grandes distances.

Quantification de la relation séquence-activité de l’ARN par prédiction de structure tridimensionnelle

Dans un premier temps, nous avons modélisé la structure d’une famille d’ARN avec une grammaire de graphes afin d’identifier les séquences qui en font partie. Plusieurs autres méthodes de modélisation ont été développées, telles que des grammaires stochastiques hors-contexte, des modèles de covariance, des profils de structures secondaires et des réseaux de contraintes. Ces méthodes de modélisation se basent sur la structure secondaire classique comparativement à nos grammaires de graphes qui se basent sur les motifs cycliques de nucléotides. Pour exemplifier notre modèle, nous avons utilisé la boucle E du ribosome qui contient le motif Sarcin-Ricin qui a été largement étudié depuis sa découverte par cristallographie aux rayons X au début des années 90. Nous avons construit une grammaire de graphes pour la structure du motif Sarcin-Ricin et avons dérivé toutes les séquences qui peuvent s’y replier. La pertinence biologique de ces séquences a été confirmée par une comparaison des séquences d’un alignement de plus de 800 séquences ribosomiques bactériennes. Cette comparaison a soulevée des alignements alternatifs pour quelques unes des séquences que nous avons supportés par des prédictions de structures secondaires et tertiaires. Les motifs cycliques de nucléotides ont été observés par les membres de notre laboratoire dans l'ARN dont la structure tertiaire a été résolue expérimentalement. Une étude des séquences et des structures tertiaires de chaque cycle composant la structure du Sarcin-Ricin a révélé que l'espace des séquences dépend grandement des interactions entre tous les nucléotides à proximité dans l’espace tridimensionnel, c’est-à-dire pas uniquement entre deux paires de bases adjacentes. Le nombre de séquences générées par la grammaire de graphes est plus petit que ceux des méthodes basées sur la structure secondaire classique. Cela suggère l’importance du contexte pour la relation entre la séquence et la structure, d’où l’utilisation d’une grammaire de graphes contextuelle plus expressive que les grammaires hors-contexte. Les grammaires de graphes que nous avons développées ne tiennent compte que de la structure tertiaire et négligent les interactions de groupes chimiques spécifiques avec des éléments extra-moléculaires, comme d’autres macromolécules ou ligands. Dans un deuxième temps et pour tenir compte de ces interactions, nous avons développé un modèle qui tient compte de la position des groupes chimiques à la surface des structures tertiaires. L’hypothèse étant que les groupes chimiques à des positions conservées dans des séquences prédéterminées actives, qui sont déplacés dans des séquences inactives pour une fonction précise, ont de plus grandes chances d’être impliqués dans des interactions avec des facteurs. En poursuivant avec l’exemple de la boucle E, nous avons cherché les groupes de cette boucle qui pourraient être impliqués dans des interactions avec des facteurs d'élongation. Une fois les groupes identifiés, on peut prédire par modélisation tridimensionnelle les séquences qui positionnent correctement ces groupes dans leurs structures tertiaires. Il existe quelques modèles pour adresser ce problème, telles que des descripteurs de molécules, des matrices d’adjacences de nucléotides et ceux basé sur la thermodynamique. Cependant, tous ces modèles utilisent une représentation trop simplifiée de la structure d’ARN, ce qui limite leur applicabilité. Nous avons appliqué notre modèle sur les structures tertiaires d’un ensemble de variants d’une séquence d’une instance du Sarcin-Ricin d’un ribosome bactérien. L’équipe de Wool à l’université de Chicago a déjà étudié cette instance expérimentalement en testant la viabilité de 12 variants. Ils ont déterminé 4 variants viables et 8 létaux. Nous avons utilisé cet ensemble de 12 séquences pour l’entraînement de notre modèle et nous avons déterminé un ensemble de propriétés essentielles à leur fonction biologique. Pour chaque variant de l’ensemble d’entraînement nous avons construit des modèles de structures tertiaires. Nous avons ensuite mesuré les charges partielles des atomes exposés sur la surface et encodé cette information dans des vecteurs. Nous avons utilisé l’analyse des composantes principales pour transformer les vecteurs en un ensemble de variables non corrélées, qu’on appelle les composantes principales. En utilisant la distance Euclidienne pondérée et l’algorithme du plus proche voisin, nous avons appliqué la technique du « Leave-One-Out Cross-Validation » pour choisir les meilleurs paramètres pour prédire l’activité d’une nouvelle séquence en la faisant correspondre à ces composantes principales. Finalement, nous avons confirmé le pouvoir prédictif du modèle à l’aide d’un nouvel ensemble de 8 variants dont la viabilité à été vérifiée expérimentalement dans notre laboratoire. En conclusion, les grammaires de graphes permettent de modéliser la relation entre la séquence et la structure d’un élément structural d’ARN, comme la boucle E contenant le motif Sarcin-Ricin du ribosome. Les applications vont de la correction à l’aide à l'alignement de séquences jusqu’au design de séquences ayant une structure prédéterminée. Nous avons également développé un modèle pour tenir compte des interactions spécifiques liées à une fonction biologique donnée, soit avec des facteurs environnants. Notre modèle est basé sur la conservation de l'exposition des groupes chimiques qui sont impliqués dans ces interactions. Ce modèle nous a permis de prédire l’activité biologique d’un ensemble de variants de la boucle E du ribosome qui se lie à des facteurs d'élongation.

Construction d'un Atlas 3D numérique de la cornée humaine par recalage d'images

Nous proposons de construire un atlas numérique 3D contenant les caractéristiques moyennes et les variabilités de la morphologie d’un organe. Nos travaux seront appliqués particulièrement à la construction d'un atlas numérique 3D de la totalité de la cornée humaine incluant la surface antérieure et postérieure à partir des cartes topographiques fournies par le topographe Orbscan II. Nous procédons tout d'abord par normalisation de toute une population de cornées. Dans cette étape, nous nous sommes basés sur l'algorithme de recalage ICP (iterative closest point) pour aligner simultanément les surfaces antérieures et postérieures d'une population de cornée vers les surfaces antérieure et postérieure d'une cornée de référence. En effet, nous avons élaboré une variante de l'algorithme ICP adapté aux images (cartes) de cornées qui tient compte de changement d'échelle pendant le recalage et qui se base sur la recherche par voisinage via la distance euclidienne pour établir la correspondance entre les points. Après, nous avons procédé pour la construction de l'atlas cornéen par le calcul des moyennes des élévations de surfaces antérieures et postérieures recalées et leurs écarts-types associés. Une population de 100 cornées saines a été utilisée pour construire l'atlas cornéen normal. Pour visualiser l’atlas, on a eu recours à des cartes topographiques couleurs similairement à ce qu’offrent déjà les systèmes topographiques actuels. Enfin, des observations ont été réalisées sur l'atlas cornéen reflétant sa précision et permettant de développer une meilleure connaissance de l’anatomie cornéenne.

Articulated Model Registration of MRI/X-Ray Spine Data

This paper presents a method based on articulated models for the registration of spine data extracted from multimodal medical images of patients with scoliosis. With the ultimate aim being the development of a complete geometrical model of the torso of a scoliotic patient, this work presents a method for the registration of vertebral column data using 3D magnetic resonance images (MRI) acquired in prone position and X-ray data acquired in standing position for five patients with scoliosis. The 3D shape of the vertebrae is estimated from both image modalities for each patient, and an articulated model is used in order to calculate intervertebral transformations required in order to align the vertebrae between both postures. Euclidean distances between anatomical landmarks are calculated in order to assess multimodal registration error. Results show a decrease in the Euclidean distance using the proposed method compared to rigid registration and more physically realistic vertebrae deformations compared to thin-plate-spline (TPS) registration thus improving alignment.

Realisation of a Target Classifier For Noise Sources in the Ocean

This thesis addresses one of the emerging topics in Sonar Signal Processing.,viz.the implementation of a target classifier for the noise sources in the ocean, as the operator assisted classification turns out to be tedious,laborious and time consuming.In the work reported in this thesis,various judiciously chosen components of the feature vector are used for realizing the newly proposed Hierarchical Target Trimming Model.The performance of the proposed classifier has been compared with the Euclidean distance and Fuzzy K-Nearest Neighbour Model classifiers and is found to have better success rates.The procedures for generating the Target Feature Record or the Feature vector from the spectral,cepstral and bispectral features have also been suggested.The Feature vector ,so generated from the noise data waveform is compared with the feature vectors available in the knowledge base and the most matching pattern is identified,for the purpose of target classification.In an attempt to improve the success rate of the Feature Vector based classifier,the proposed system has been augmented with the HMM based Classifier.Institutions where both the classifier decisions disagree,a contention resolving mechanism built around the DUET algorithm has been suggested.

Cashew industry in Kerala problems and potentials

India is the largest producer and processor of cashew in the world. The export value of cashew is about Rupees 2600 crore during 2004-05. Kerala is the main processing and exporting center of cashew. In Kerala most of the cashew processing factories are located in Kollam district. The industry provides livelihood for about 6-7 lakhs of employees and farmers, the cashew industry has national importance. In Kollam district alone there are more than 2.5 lakhs employees directly involved in the industry, which comes about 10 per cent of the population of the district, out of which 95 per cent are women workers. It is a fact that any amount received by a woman worker will be utilized directly for the benefit of the family and hence the link relating to family welfare is quite clear. Even though the Government of Kerala has incorporated the Kerala State Cashew Development Corporation (KSCDC) and Kerala State Cashew Workers Apex Industrial Co—operative Society (CAPEX) to develop the Cashew industry, the cashew industry and ancillary industries did not grow as per the expectation. In this context, an attempt has been made to analyze the problems and potential of the industry so as to make the industry viable and sustainable for the perpetual employment and income generation as well as the overall development of the Kollam district.

Speech sample estimation from composite zerocrossings and encoding via adaptive switching of transforms

This thesis investigates the potential use of zerocrossing information for speech sample estimation. It provides 21 new method tn) estimate speech samples using composite zerocrossings. A simple linear interpolation technique is developed for this purpose. By using this method the A/D converter can be avoided in a speech coder. The newly proposed zerocrossing sampling theory is supported with results of computer simulations using real speech data. The thesis also presents two methods for voiced/ unvoiced classification. One of these methods is based on a distance measure which is a function of short time zerocrossing rate and short time energy of the signal. The other one is based on the attractor dimension and entropy of the signal. Among these two methods the first one is simple and reguires only very few computations compared to the other. This method is used imtea later chapter to design an enhanced Adaptive Transform Coder. The later part of the thesis addresses a few problems in Adaptive Transform Coding and presents an improved ATC. Transform coefficient with maximum amplitude is considered as ‘side information’. This. enables more accurate tfiiz assignment enui step—size computation. A new bit reassignment scheme is also introduced in this work. Finally, sum ATC which applies switching between luiscrete Cosine Transform and Discrete Walsh-Hadamard Transform for voiced and unvoiced speech segments respectively is presented. Simulation results are provided to show the improved performance of the coder

CBIR Using Local and Global Properties of Image Sub-blocks

This paper proposes a content based image retrieval (CBIR) system using the local colour and texture features of selected image sub-blocks and global colour and shape features of the image. The image sub-blocks are roughly identified by segmenting the image into partitions of different configuration, finding the edge density in each partition using edge thresholding, morphological dilation and finding the corner density in each partition. The colour and texture features of the identified regions are computed from the histograms of the quantized HSV colour space and Gray Level Co- occurrence Matrix (GLCM) respectively. A combined colour and texture feature vector is computed for each region. The shape features are computed from the Edge Histogram Descriptor (EHD). Euclidean distance measure is used for computing the distance between the features of the query and target image. Experimental results show that the proposed method provides better retrieving result than retrieval using some of the existing methods

Content Based Image Retrieval Using Low Level Features of Automatically Extracted Regions of Interest

This paper proposes a region based image retrieval system using the local colour and texture features of image sub regions. The regions of interest (ROI) are roughly identified by segmenting the image into fixed partitions, finding the edge map and applying morphological dilation. The colour and texture features of the ROIs are computed from the histograms of the quantized HSV colour space and Gray Level co- occurrence matrix (GLCM) respectively. Each ROI of the query image is compared with same number of ROIs of the target image that are arranged in the descending order of white pixel density in the regions, using Euclidean distance measure for similarity computation. Preliminary experimental results show that the proposed method provides better retrieving result than retrieval using some of the existing methods.