Abstract This thesis presents a multiscale framework for the construction of affine invariant pattern recognition and registration methods. The idea in the introduced approach is to extend the given pattern to a set of affine covariant versions, each carrying slightly different information, and then to apply known affine invariants to each of them separately. The key part of the framework is the construction of the affine covariant set, and this is done by combining several scaled representations of the original pattern. The advantages compared to previous approaches include the possibility of many variations and the inclusion of spatial information on the patterns in the features. The application of the multiscale framework is demonstrated by constructing several new affine invariant methods using different preprocessing techniques, combination schemes, and final recognition and registration approaches. The techniques introduced are briefly described from the perspective of the multiscale framework, and further treatment and properties are presented in the corresponding original publications. The theoretical discussion is supported by several experiments where the new methods are compared to existing approaches. In this thesis the patterns are assumed to be gray scale images, since this is the main application where affine relations arise. Nevertheless, multiscale methods can also be applied to other kinds of patterns where an affine relation is present. An additional application of one multiscale based technique in convexity measurements is introduced. The method, called multiscale autoconvolution, can be used to build a convexity measure which is a descriptor of object shape. The proposed measure has two special features compared to existing approaches. It can be applied directly to gray scale images approximating binary objects, and it can be easily modified to produce a number of measures. The new measure is shown to be straightforward to evaluate for a given shape, and it performs well in the applications, as demonstrated by the experiments in the original paper.

The Department of Arequipa, in Peru, is a region with limited water resources making freshwater management critical and requiring the development of water-demand models, which can be valuable tools for policymakers. This study developed a monthly agricultural water-demand mapping algorithm for the agricultural districts surrounding the city of Arequipa. It was accomplished by:(1) developing a ground-reference data collection method;(2) creating a crop mapping algorithm, which incorporates supervised classification methods, as well as spatial-and temporal-consistency correction methods to create crop maps out of high resolution (~3 m) PlanetScope satellite images; (3) integrating a crop growth-stage prediction algorithm for the crop maps and; (4) applying an algorithm for the estimation of the agricultural-water-demand maps using the results of steps 2 and 3, local climate data, and an irrigation demand estimation tool. The crop mapping algorithm was shown to create maps with acceptable accuracy, with 5 out of 6 months with available data having mean monthly classification accuracies of 69% to 77%for those classes which had available data. Growth stage predictions had mean absolute prediction errors of 0.55 to 0.69 months in 5 out of 6 months.The6th month (the first with ground reference data collection) had a mean absolute prediction error of 0.90 months because it lacked prior month information to correctly identify planting month. Water demand maps were produced with high spatial (3.0m) and temporal (monthly) resolution, allowing for a detailed look at local agricultural water demands. This study provides a framework for future large-scale agricultural-water demand mapping for the Department of Arequipa and similar regions around the world.

No description available.

No / Triple negative breast cancers (TNBC) are clinically heterogeneous, an aggressive subtype with poor diagnosis and strong resistance to therapy. There is a need to identify novel robust biomarkers with high specificity for early detection and therapeutic intervention. Microarray gene expression-based studies have offered significant advances in molecular classification and identification of diagnostic/prognostic signatures, however sample scarcity and cohort heterogeneity remains area of concern. In this study, we performed integrated analysis on independent microarray big data studies and identified a robust 880-gene signature for TNBC diagnosis. We further identified 16-gene (OGN, ESR1, GPC3, LHFP, AGR3, LPAR1, LRRC17, TCEAL1, CIRBP, NTN4, TUBA1C, TMSB10, RPL27, RPS3A, RPS18, and NOSTRIN) that are associated to TNBC tissues. The 880-gene signature achieved excellent classification accuracy ratio on each independent expression data sets with overall average of 99.06%, is an indication of its diagnostic power. Gene ontology enrichment analysis of 880-gene signature shows that cell-cycle pathways/processes are important clinical targets for triple negative breast cancer. Further verification of 880-gene signature could provide additive knowledge for better understanding and future direction of triple negative breast cancer research.

Approved for public release; distribution in unlimited. / In recent years there has been an increased interest in effective individual control and enhanced security measures, and face recognition schemes play an important role in this increasing market. In the past, most face recognition research studies have been conducted with visible imaging data. Only recently have IR imaging face recognition studies been reported for wide use applications, as uncooled IR imaging technology has improved to the point where the resolution of these much cheaper cameras closely approaches that of cooled counterparts. This study is part of an on-going research conducted at the Naval Postgraduate School which investigates the feasibility of applying a low cost uncooled IR camera for face recognition applications. This specific study investigates whether nonlinear kernel-based classifiers may improve overall classification rates over those obtained with linear classification schemes. The study is applied to a 50 subject IR database developed in house with a low resolution uncooled IR camera. Results show best overall mean classification performances around 98.55% which represents a 5% performance improvement over the best linear classifier results obtained previously on the same database. This study also considers several metrics to evaluate the impacts variations in various user-specified parameters have on the resulting classification performances. These results show that a low-cost, low-resolution IR camera combined with an efficient classifier can play an effective role in security related applications. / Captain, Hellenic Air Force

Class imbalanced datasets constitute a significant portion of the machine learning problems of interest, where recog­nizing the ‘rare class’ is the primary objective for most applications. Traditional linear machine learning algorithms are often not effective in recognizing the rare class. In this research work, a specifically optimized feed-forward artificial neural network (ANN) is proposed and developed to train from moderate to highly imbalanced datasets. The proposed methodology deals with the difficulty in classification task in multiple stages—by optimizing the training dataset, modifying kernel function to generate the gram matrix and optimizing the NN structure. First, the training dataset is extracted from the available sample set through an iterative process of selective under-sampling. Then, the proposed artificial NN comprises of a kernel function optimizer to specifically enhance class boundaries for imbalanced datasets by conformally transforming the kernel functions. Finally, a single hidden layer weighted neural network structure is proposed to train models from the imbalanced dataset. The proposed NN architecture is derived to effectively classify any binary dataset with even very high imbalance ratio with appropriate parameter tuning and sufficient number of processing elements. Effectiveness of the proposed method is tested on accuracy based performance metrics, achieving close to and above 90%, with several imbalanced datasets of generic nature and compared with state of the art methods. The proposed model is also used for classification of a 25GB computed tomographic colonography database to test its applicability for big data. Also the effectiveness of under-sampling, kernel optimization for training of the NN model from the modified kernel gram matrix representing the imbalanced data distribution is analyzed experimentally. Computation time analysis shows the feasibility of the system for practical purposes. This report is concluded with discussion of prospect of the developed model and suggestion for further development works in this direction.

Nature has created complex biological structures that exhibit intelligent behaviour through an evolutionary process. Thus, intelligence and evolution are intimately connected. This has inspired evolutionary computation (EC) that simulates the evolutionary process to develop powerful techniques such as genetic algorithms (GAs), genetic programming (GP), evolutionary strategies (ES) and evolutionary programming (EP) to solve real-world problems in learning, control, optimization and classification. GP discovers the relationship among data and expresses it as a LISP-S expression i.e., a computer program. Thus the goal of program discovery as a solution for a problem is addressed by GP in the framework of evolutionary computation. In this thesis, we address for the first time the problem of applying GP to mu1ticategory pattern classification. In supervised pattern classification, an input vector of m dimensions is mapped onto one of the n classes. It has a number of application areas such as remote sensing, medical diagnosis etc., A supervised classifier is developed by using a training set that contains representative samples of various classes present in the application. Supervised classification has been done earlier with maximum likelihood classifier: neural networks and fuzzy logic. The major considerations in applying GP to pattern classification are listed below: (i) GP-based techniques are data distribution-free i.e., no a priori knowledge is needed abut the statistical distribution of the data or no assumption such as normal distribution for data needs to be made as in MLC. (ii) GP can directly operate on the data in its original form. (iii) GP can detect the underlying but unknown relationship that mists among data and express it as a mathematical LISP S-expression. The generated LISP S-expressions can be directly used in the application environment. (iv) GP can either discover the most important discriminating features of a class during evolution or it requires minor post-processing of the LISP-S expression to discover the discriminant features. In a neural network, the knowledge learned by the neural network about the data distributions is embedded in the interconnection weights and it requires considerable amount of post-processing of the weights to understand the decision of the neural network. In 2-category pattern classification, a single GP expression is evolved as a discriminant function. The output of the GP expression can be +l for samples of one class and -1 for samples of the other class. When the GP paradigm is applied to an n-class problem, the following questions arise: Ql. As a typical GP expression returns a value (+l or -1) for a 2-class problem, how does one apply GP for the n-class pattern classification problem? Q2. What should be the fitness function during evolution of the GP expressions? Q3. How does the choice of a function set affect the performance of GP-based classification? Q4. How should training sets be created for evaluating fitness during the evolution of GP classifier expressions? Q5. How does one improve learning of the underlying data distributions in a GP framework? Q6. How should conflict resolution be handled before assigning a class to the input feature vector? Q7. How does GP compare with other classifiers for an n-class pattern classification problem? The research described here seeks to answer these questions. We show that GP can be applied to an n-category pattern classification problem by considering it as n 2-class problems. The suitability of this approach is demonstrated by considering a real-world problem based on remotely sensed satellite images and Fisher's Iris data set. In a 2-class problem, simple thresholding is sufficient for a discriminant function to divide the feature space into two regions. This means that one genetic programming classifier expression (GPCE) is sufficient to say whether or not the given input feature vector belongs to that class; i.e., the GP expression returns a value (+1 or -1). As the n-class problem is formulated as n 2-class problems, n GPCEs are evolved. Hence, n GPCE specific training sets are needed to evolve these n GPCEs. For the sake of illustration, consider a 5-class pat tern classification problem. Let n, be the number of samples that belong to class j, and N, be the number of samples that do not belong to class j, (j = 1,..., 5). Thus, N1=n2+n3+n4+n5 N2=n1+n3+n4+n5 N3=n1+n2+n4+n5 N4=n1+n2+n3+n5 N5=n1+n2+n3+n4 Thus, When the five class problem is formulated as five 2-class problems. we need five GPCEs as discriminant functions to resolve between n1 and N1, n2 and N2, n3 and N3, n4 and N4 and lastly n5 and N5. Each of these five 2-class problems is handled as a separate 2-class problem with simple thresholding. Thus, GPCE# l resolves between samples of class# l and the remaining n - 1 classes. A training set is needed to evaluate the fitness of GPCE during its evolution. If we directly create the training set, it leads to skewness (as n1 < N1). To overcome the skewness, an interleaved data format is proposed for the training set of a GPCE. For example, in the training set of GPCE# l, samples of class# l are placed alternately between samples of the remaining n - 1 classes. Thus, the interleaved data format is an artifact to create a balanced training set. Conventionally, all the samples of a training set are fed to evaluate the fitness of every member of the population in each generation. We call this "global" learning 3s GP tries to learn the entire training set at every stage of the evolution. We have introduced incremental learning to simplify the task of learning for the GP paradigm. A subset of the training set is fed and the size of the subset is gradually increased over time to cover the entire training data. The basic motivation for incremental learning is to improve learning during evolution as it is easier to learn a smaller task and then to progress from a smaller task to a bigger task. Experimental results are presented to show that the interleaved data format and incremental learning improve the performance of the GP classifier. We also show that the GPCEs evolved with an arithmetic function set are able to track variation in the input better than GPCEs evolved with function sets containing logical and nonlinear elements. Hence, we have used arithmetic function set, incremental learning, and interleaved data format to evolve GPCEs in our simulations. AS each GPCE is trained to recognize samples belonging to its own class and reject samples belonging to other classes a strength of association measure is associated with each GPCE to indicate the degree to which it can recognize samples belonging to its own class. The strength of association measures are used for assigning a class to an input feature vector. To reduce misclassification of samples, we also show how heuristic rules can be generated in the GP framework unlike in either MLC or the neural network classifier. We have also studied the scalability and generalizing ability of the GP classifier by varying the number of classes. We also analyse the performance of the GP classifier by considering the well-known Iris data set. We compare the performance of classification rules generated from the GP classifier with those generated from neural network classifier, (24.5 method and fuzzy classifier for the Iris data set. We show that the performance of GP is comparable to other classifiers for the Iris data set. We notice that the classification rules can be generated with very little post-processing and they are very similar to the rules generated from the neural network and C4.5 for the Iris data set. Incremental learning influences the number of generations available for GP to learn the data distribution of classes whose d is -1 in the interleaved data format. This is because the samples belonging to the true class (desired output d is +1) are alternately placed between samples belonging to other classes i.e., they are repeated to balance the training set in the interleaved data format. For example, in the evolution of GPCE for class# l, the fitness function can be fed initially with samples of class#:! and subsequently with the samples of class#3, class#4 and class#. So in the evaluation of the fitness function, the samples of class#kt5 will not be present when the samples of class#2 are present in the initial stages. However, in the later stages of evolution, when samples of class#5 are fed, the fitness function will utilize the samples of both class#2 and class#5. As learning in evolutionary computation is guided by the evaluation of the fitness function, GPCE# l gets lesser number of generations to learn how to reject data of class#5 as compared to the data of class#2. This is because the termination criterion (i.e., the maximum number of generations) is defined a priori. It is clear that there are (n-l)! Ways of ordering the samples of classes whose d is -1 in the interleaved data format. Hence a heuristic is presented to determine a possible order to feed data of different classes for the GPCEs evolved with incremental learning and interleaved data format. The heuristic computes an overlap index for each class based on its spatial spread and distribution of data in the region of overlap with respect to other classes in each feature. The heuristic determines the order in which classes whose desired output d is –1 should be placed in each GPCE-specific training set for the interleaved data format. This ensures that GP gets more number of generations to learn about the data distribution of a class with higher overlap index than a class with lower overlap index. The ability of the GP classifier to learn the data distributions depends upon the number of classes and the spatial spread of data. As the number of classes increases, the GP classifier finds it difficult to resolve between classes. So there is a need to partition the feature space and identify subspaces with reduced number of classes. The basic objective is to divide the feature space into subspaces and hence the data set that contains representative samples of n classes into subdata sets corresponding to the subspaces of the feature space, so that some of the subdata sets/spaces can have data belonging to only p classes (p < n). The GP classifier is then evolved independently for the subdata sets/spaces of the feature space. This results in localized learning as the GP classifier has to learn the data distribution in only a subspace of the feature space rather than in the entire feature space. By integrating the GP classifier with feature space partitioning (FSP), we improve classification accuracy due to localized learning. Although serial computers have increased steadily in their performance, the quest for parallel implementation of a given task has continued to be of interest in any computationally intensive task since parallel implementation leads to a faster execution than a serial implementation As fitness evaluation, selection strategy and population structures are used to evolve a solution in GP, there is scope for a parallel implementation of GP classifier. We have studied distributed GP and massively parallel GP for our approach to GP-based multicategory pattern classification. We present experimental results for distributed GP with Message Passing Interface on IBM SP2 to highlight the speedup that can be achieved over the serial implementation of GP. We also show how data parallelism can be used to further speed up fitness evaluation and hence the execution of the GP paradigm for multicategory pat tern classification. We conclude that GP can be applied to n-category pattern classification and its potential lies in its simplicity and scope for parallel implementation. The GP classifier developed in this thesis can be looked upon as an addition to the earlier statistical, neural and fuzzy approaches to multicategory pattern classification.

The growth in the use of machine learning in different areas has revealed challenging classification problems that require robust systems. Multiple Classier Systems (MCSs) have attracted interest from researchers as a method that could address such problems. Optimizing the fitness of an MCS improves its, robustness. The lack of an analysis for MCSs from a fitness perspective is identified. To fill this gap, an analytic model from this perspective is derived mathematically by extending the error analysis introduced by Brown and Kuncheva in 2010. The model relates the fitness of an MCS to the average accuracy, positive-diversity, and negative-diversity of the classifiers that constitute the MCS. The model is verified using a statistical analysis of a Monte-Carlo based simulation. This shows the significance of the indicated relationships by the model. This model provides guidelines for developing robust MCSs. It enables the selection of classifiers which compose an MCS with an improved fitness while improving computational cost by avoiding local calculations. The usefulness of the model for designing classification systems is investigated. A new measure consisting of the accuracy and positive-diversity is developed. This measure evaluates fitness while avoiding many calculations compared to the regular measures. A new system (Gadapt) is developed. Gadapt combines machine learning and genetic algorithms to define subsets of the feature space that closely match true class regions. It uses the new measure as a multi-objective criterion for a multi-objective genetic algorithm to identify the MCSs those create the subsets. The design of Gadapt is validated experimentally. The usefulness of the measure and the method of determining the subsets for the performance of Gadapt are examined based on five generated data sets that represent a wide range of problems. The robustness of Gadapt to small amounts of training data is evaluated in comparison with five existing systems on four benchmark data sets. The performance of Gadapt is evaluated in comparison with eleven existing systems on nine benchmark data sets. The analysis of the experiment results supports the validity of the Gadapt design and the outperforming of Gadapt on the existing systems in terms of robustness and performance.

The aim of this thesis is to develop a stereo vision system to locate and classify objects moving on a conveyor belt. The vision system determines the locations of the objects with respect to a world coordinate system and class of the objects. In order to estimate the locations of the objects, two cameras placed at different locations are used. Image processing algorithms are employed to extract some features of the objects. These features are fed to stereo matching and classifier algorithms. The results of stereo matching algorithm are combined with the calibration parameters of the cameras to determine the object locations. Pattern classification techniques (Bayes and Nearest Neighbor classifiers) are used to classify the objects. The linear velocity of the objects is determined by using an encoder mounted to the shaft of the motor driving the conveyor belt. A robot can plan a sequence of motion to pick the object from the conveyor belt by using the output of the proposed system.

Three-dimensional movement analysis was used to evaluate head and neck movement in patients with neck pain and matched controls. The aims were to further develop biomechanical models of head and neck kinematics, to investigate differences between subjects with non-specific neck pain and whiplash associated disorders (WAD), and to evaluate the potential of objective movement analysis as a decision support during diagnosis and follow-up of patients with neck pain. Fast, repetitive head movements (flexion, extension, rotation to the side) were studied in a group of 59 subjects with WAD and 56 controls. A back propagation artificial neural network classified vectors of collected movement variables from each individual according to group membership with a predictivity of 89%. The helical axis for head movement were analyzed in two groups of neck pain patients (21 with non-specific neck pain and 22 with WAD) and 24 matched controls. A moving time window with a cut-off angle of 4° was used to calculate finite helical axes. The centre of rotation of the finite axes (CR) was derived as the 3D intersection point of the finite axes. A downward migration of the axis during flexion/extension and a change of axis direction towards the end of the movements were observed. CR was at its most superior position during side rotations and at its most inferior during ball catching. This could relate to that side rotation was mainly done in the upper spine, while all cervical vertebrae were recruited to stabilize the head in the more complex catching task. Changes in movement strategy were observed in the neck pain groups: Neck pain subjects had lower mean velocities and ranges of movements as compared with controls during ball catching, which could relate to a stiffer body position in neck pain patients in order to stabilize the neck. In addition, the WAD group had a displaced axis position during head repositioning after flexion, while CR was displaced during fast side rotations in the non-specific neck pain group. Pain intensity correlated with axis and CR position, and may be one reason for the movement strategy changes. Increased amount of irregularities in the trajectory of the axis was found in the WAD group during head repositioning, fast repetitive head movements and catching. This together with an increased constant repositioning error during repositioning after flexion indicated motor control disturbances. A higher group standard deviation in neck pain groups indicated heterogeneity among subjects in this disturbance. Wireless motion sensors and electro-oculography was used simultaneously, as an initial step towards a portable system and towards a method to quantify head-eye co-ordination deficits in individuals with WAD. Twenty asymptomatic control subjects and six WAD subjects with eye disturbances (e.g. dizziness and double vision) were studied. The trial-to-trial repeatability was moderate to high for all evaluated variables (single intraclass correlation coefficients &gt;0.4 in 28 of 32 variables). The WAD subjects demonstrated decreased head velocity, decreased range of head movement during gaze fixation and lowered head stability during head-eye co-ordination as possible deficits. In conclusion, kinematical analyses have a potential to be used as a support for physicians and physiotherapists for diagnosis and follow-up of neck pain patients. Specifically, the helical axis method gives information about how the movement is performed. However, a flexible motion capture system (for example based on wireless motion sensors) is needed. Combined analysis of several variables is preferable, as patients with different neck pain disorders seem to be a heterogeneous group.