The overarching goal of this project was to improve breast cancer screening protocols first by collecting small angle x-ray scattering (SAXS) images from breast biopsy tissue, and second, by applying pattern recognition techniques as a semi-automatic screen. Wavelet based features were generated from the SAXS image data. The features were supplied to a classifier, which sorted the images into distinct groups, such as normal and tumor. <p>The main problem in the project was to find a set of features that provided sufficient separation for classification into groups of normal and tumor. In the original SAXS patterns, information useful for classification was obscured. The wavelet maps allowed new scale-based information to be uncovered from each SAXS pattern. The new information was subsequently used to define features that allowed for classification. Several calculations were tested to extract useful features from the wavelet decomposition maps. The wavelet map average intensity feature was selected as the most promising feature. The wavelet map intensity feature was improved by using pre-processing to remove the high central intensities from the SAXS patterns, and by using different wavelet bases for the wavelet decomposition. <p>The investigation undertaken for this project showed very promising results. A classification rate of 100% was achieved for distinguishing between normal samples and tumor samples. The system also showed promising results when tested on unrelated MRI data. In the future, the semi-automatic pattern recognition tool developed for this project could be automated. With a larger set of data for training and testing, the tool could be improved upon and used to assist radiologists in the detection and classification of breast lesions.

EMG signal decomposition is the process of resolving a composite EMG signal into its constituent motor unit potential trains (classes) and it can be configured as a classification problem. An EMG signal detected by the tip of an inserted needle electrode is the superposition of the individual electrical contributions of the different motor units that are active, during a muscle contraction, and background interference. <BR>This thesis addresses the process of EMG signal decomposition by developing an interactive classification system, which uses multiple classifier fusion techniques in order to achieve improved classification performance. The developed system combines heterogeneous sets of base classifier ensembles of different kinds and employs either a one level classifier fusion scheme or a hybrid classifier fusion approach. <BR>The hybrid classifier fusion approach is applied as a two-stage combination process that uses a new aggregator module which consists of two combiners: the first at the abstract level of classifier fusion and the other at the measurement level of classifier fusion such that it uses both combiners in a complementary manner. Both combiners may be either data independent or the first combiner data independent and the second data dependent. For the purpose of experimentation, we used as first combiner the majority voting scheme, while we used as the second combiner one of the fixed combination rules behaving as a data independent combiner or the fuzzy integral with the lambda-fuzzy measure as an implicit data dependent combiner. <BR>Once the set of motor unit potential trains are generated by the classifier fusion system, the firing pattern consistency statistics for each train are calculated to detect classification errors in an adaptive fashion. This firing pattern analysis allows the algorithm to modify the threshold of assertion required for assignment of a motor unit potential classification individually for each train based on an expectation of erroneous assignments. <BR>The classifier ensembles consist of a set of different versions of the Certainty classifier, a set of classifiers based on the nearest neighbour decision rule: the fuzzy <em>k</em>-NN and the adaptive fuzzy <em>k</em>-NN classifiers, and a set of classifiers that use a correlation measure as an estimation of the degree of similarity between a pattern and a class template: the matched template filter classifiers and its adaptive counterpart. The base classifiers, besides being of different kinds, utilize different types of features and their performances were investigated using both real and simulated EMG signals of different complexities. The feature sets extracted include time-domain data, first- and second-order discrete derivative data, and wavelet-domain data. <BR>Following the so-called <em>overproduce and choose</em> strategy to classifier ensemble combination, the developed system allows the construction of a large set of candidate base classifiers and then chooses, from the base classifiers pool, subsets of specified number of classifiers to form candidate classifier ensembles. The system then selects the classifier ensemble having the maximum degree of agreement by exploiting a diversity measure for designing classifier teams. The kappa statistic is used as the diversity measure to estimate the level of agreement between the base classifier outputs, i. e. , to measure the degree of decision similarity between the base classifiers. This mechanism of choosing the team's classifiers based on assessing the classifier agreement throughout all the trains and the unassigned category is applied during the one level classifier fusion scheme and the first combiner in the hybrid classifier fusion approach. For the second combiner in the hybrid classifier fusion approach, we choose team classifiers also based on kappa statistics but by assessing the classifiers agreement only across the unassigned category and choose those base classifiers having the minimum agreement. <BR>Performance of the developed classifier fusion system, in both of its variants, i. e. , the one level scheme and the hybrid approach was evaluated using synthetic simulated signals of known properties and real signals and then compared it with the performance of the constituent base classifiers. Across the EMG signal data sets used, the hybrid approach had better average classification performance overall, specially in terms of reducing the number of classification errors.

EMG signal decomposition is the process of resolving a composite EMG signal into its constituent motor unit potential trains (classes) and it can be configured as a classification problem. An EMG signal detected by the tip of an inserted needle electrode is the superposition of the individual electrical contributions of the different motor units that are active, during a muscle contraction, and background interference. <BR>This thesis addresses the process of EMG signal decomposition by developing an interactive classification system, which uses multiple classifier fusion techniques in order to achieve improved classification performance. The developed system combines heterogeneous sets of base classifier ensembles of different kinds and employs either a one level classifier fusion scheme or a hybrid classifier fusion approach. <BR>The hybrid classifier fusion approach is applied as a two-stage combination process that uses a new aggregator module which consists of two combiners: the first at the abstract level of classifier fusion and the other at the measurement level of classifier fusion such that it uses both combiners in a complementary manner. Both combiners may be either data independent or the first combiner data independent and the second data dependent. For the purpose of experimentation, we used as first combiner the majority voting scheme, while we used as the second combiner one of the fixed combination rules behaving as a data independent combiner or the fuzzy integral with the lambda-fuzzy measure as an implicit data dependent combiner. <BR>Once the set of motor unit potential trains are generated by the classifier fusion system, the firing pattern consistency statistics for each train are calculated to detect classification errors in an adaptive fashion. This firing pattern analysis allows the algorithm to modify the threshold of assertion required for assignment of a motor unit potential classification individually for each train based on an expectation of erroneous assignments. <BR>The classifier ensembles consist of a set of different versions of the Certainty classifier, a set of classifiers based on the nearest neighbour decision rule: the fuzzy <em>k</em>-NN and the adaptive fuzzy <em>k</em>-NN classifiers, and a set of classifiers that use a correlation measure as an estimation of the degree of similarity between a pattern and a class template: the matched template filter classifiers and its adaptive counterpart. The base classifiers, besides being of different kinds, utilize different types of features and their performances were investigated using both real and simulated EMG signals of different complexities. The feature sets extracted include time-domain data, first- and second-order discrete derivative data, and wavelet-domain data. <BR>Following the so-called <em>overproduce and choose</em> strategy to classifier ensemble combination, the developed system allows the construction of a large set of candidate base classifiers and then chooses, from the base classifiers pool, subsets of specified number of classifiers to form candidate classifier ensembles. The system then selects the classifier ensemble having the maximum degree of agreement by exploiting a diversity measure for designing classifier teams. The kappa statistic is used as the diversity measure to estimate the level of agreement between the base classifier outputs, i. e. , to measure the degree of decision similarity between the base classifiers. This mechanism of choosing the team's classifiers based on assessing the classifier agreement throughout all the trains and the unassigned category is applied during the one level classifier fusion scheme and the first combiner in the hybrid classifier fusion approach. For the second combiner in the hybrid classifier fusion approach, we choose team classifiers also based on kappa statistics but by assessing the classifiers agreement only across the unassigned category and choose those base classifiers having the minimum agreement. <BR>Performance of the developed classifier fusion system, in both of its variants, i. e. , the one level scheme and the hybrid approach was evaluated using synthetic simulated signals of known properties and real signals and then compared it with the performance of the constituent base classifiers. Across the EMG signal data sets used, the hybrid approach had better average classification performance overall, specially in terms of reducing the number of classification errors.

Phoneme classification is a key area of speech recognition. Phonemes are the basic modeling units in modern speech recognition and they are the constructive units of words. Thus, being able to quickly and accurately classify phonemes that are input to a speech-recognition system is a basic and important step towards improving and eventually perfecting speech recognition as a whole. Many classification approaches currently exist that can be applied to the task of classifying phonemes. These techniques range from simple ones such as the nearest centroid classifier to complex ones such as support vector machine. Amongst the existing classifiers, the simpler ones tend to be quicker to train but have lower accuracy, whereas the more complex ones tend to be higher in accuracy but are slower to train. Because phoneme classification involves very large datasets, it is desirable to have classifiers that are both quick to train and are high in accuracy. The formulation of such classifiers is still an active ongoing research topic in phoneme classification. One paradigm in formulating such classifiers attempts to increase the accuracies of the simpler classifiers with minimal sacrifice to their running times. The opposite paradigm attempts to increase the training speeds of the more complex classifiers with minimal sacrifice to their accuracies. The objective of this research is to develop a new centroid-based classifier that builds upon the simpler nearest centroid classifier by incorporating a new discriminative locally-adaptive training procedure developed from recent advances in machine learning. This new classifier, which is referred to as the discriminative locally-adaptive nearest centroid (DLANC) classifier, achieves much higher accuracies as compared to the nearest centroid classifier whilst having a relatively low computational complexity and being able to scale up to very large datasets.

The overarching goal of this project was to improve breast cancer screening protocols first by collecting small angle x-ray scattering (SAXS) images from breast biopsy tissue, and second, by applying pattern recognition techniques as a semi-automatic screen. Wavelet based features were generated from the SAXS image data. The features were supplied to a classifier, which sorted the images into distinct groups, such as normal and tumor. <p>The main problem in the project was to find a set of features that provided sufficient separation for classification into groups of normal and tumor. In the original SAXS patterns, information useful for classification was obscured. The wavelet maps allowed new scale-based information to be uncovered from each SAXS pattern. The new information was subsequently used to define features that allowed for classification. Several calculations were tested to extract useful features from the wavelet decomposition maps. The wavelet map average intensity feature was selected as the most promising feature. The wavelet map intensity feature was improved by using pre-processing to remove the high central intensities from the SAXS patterns, and by using different wavelet bases for the wavelet decomposition. <p>The investigation undertaken for this project showed very promising results. A classification rate of 100% was achieved for distinguishing between normal samples and tumor samples. The system also showed promising results when tested on unrelated MRI data. In the future, the semi-automatic pattern recognition tool developed for this project could be automated. With a larger set of data for training and testing, the tool could be improved upon and used to assist radiologists in the detection and classification of breast lesions.

The electronic mail has become one of the most popular communication channels in the modern world. Due to its convenience and low cost, however, many business salesmen utilize this channel to promote their products by distributing e-mails to people as far as they can reach, which causes troubles to irrelevant e-mail receivers. As a result, many a research has been devoted to filtering irrelevant e-mails based on data mining techniques to alleviate users¡¦ mental loadings in processing e-mails they receive. Nevertheless, current approaches have their own drawbacks. Issues on what appropriate classifies to construct, how to endow such classifiers with the adaptive learning ability, and how to customize the e-mail management process for each user are still under investigation. The objective of this research is therefore to construct an e-mail classifier with learning ability to self-correct from erroneous outcomes. Furthermore, we propose a customized e-mail management process that can handle users¡¦ e-mails based on their own preferences. Ultimately, it can adapt itself to the changes of users¡¦ preferences when handling their e-mails. Several experiments are conducted to verify the performance of the constructed classifier. The results show that our proposed classifier possesses high accuracy and high precision with outstanding adaptive learning ability. We also illustrate a real application of the customized e-mail management process. It shows that our approach can detect the changes of users¡¦ preferences and learn to follow the changes. The feasibility of employing our approach to constructing e-mail classifiers is thus justified.

Given a limited number of samples for classification, several issues arise with respect to design, performance and analysis of classifiers. This is especially so in the case of microarray-based classification. In this paper, we use a complexity measure based mixture model to study classifier performance for small sample problems. The motivation behind such a study is to determine the conditions under which a certain class of classifiers is suitable for classification, subject to the constraint of a limited number of samples being available. Classifier study in terms of the VC dimension of a learning machine is also discussed.

In a time when cattle herds grow continually larger the need for automatic methods to detect diseases is ever increasing. One possible method to discover diseases is to use thermal images and automatic head and eye detectors. In this thesis an eye detector and a head detector is implemented using the Random Forests classifier. During the implementation the classifier is evaluated using three different descriptors: Histogram of Oriented Gradients, Local Binary Patterns, and a descriptor based on pixel differences. An alternative classifier, the Support Vector Machine, is also evaluated for comparison against Random Forests. The thesis results show that Histogram of Oriented Gradients performs well as a description of cattle heads, while Local Binary Patterns performs well as a description of cattle eyes. The provided descriptor performs almost equally well in both cases. The results also show that Random Forests performs approximately as good as the Support Vector Machine, when the Support Vector Machine is paired with Local Binary Patterns for both heads and eyes. Finally the thesis results indicate that it is easier to detect and locate cattle heads than it is to detect and locate cattle eyes. For eyes, combining a head detector and an eye detector is shown to give a better result than only using an eye detector. In this combination heads are first detected in images, followed by using the eye detector in areas classified as heads.

There exist many data classification methods and algorithms; however the importance of them has not diminished. The data and information quantities increase as well as the diversity of information, so the question is how to reliably process the data. Various considerations emerge concerning what method to choose or which of them fits for data best, i.e. which of them would classify data most accurately and reliably. This work presents a classification method based on the statistics of Dirichlet mixtures. Dirichlet mixture combines more than one Dirichlet densities that are described by the same set of parameters but with different values of them. Such a Dirichlet mixture becomes sensitive to recognize differently distributed variables (data) and hence, utilization of the Dirichlet mixtures in classification can provide a powerful tool for classification of data of any kind. This thesis proposes a method describing how Dirichlet mixtures can be utilized for classification of data of any kind. With regard to this, a Dirichlet mixtures classifier is designed to classify data of any type and with any range of values. The designed classifier classifies numerical data as well as symbolic ones. The Dirichlet mixtures classifier is implemented in two ways: The first one concerns the classifier as the end-product for a user and the second one relates to a compiled library of classification routines. Using the Dirichlet mixtures classifier as the product, the user can classify data and... [to full text]

Phoneme classification is a key area of speech recognition. Phonemes are the basic modeling units in modern speech recognition and they are the constructive units of words. Thus, being able to quickly and accurately classify phonemes that are input to a speech-recognition system is a basic and important step towards improving and eventually perfecting speech recognition as a whole. Many classification approaches currently exist that can be applied to the task of classifying phonemes. These techniques range from simple ones such as the nearest centroid classifier to complex ones such as support vector machine. Amongst the existing classifiers, the simpler ones tend to be quicker to train but have lower accuracy, whereas the more complex ones tend to be higher in accuracy but are slower to train. Because phoneme classification involves very large datasets, it is desirable to have classifiers that are both quick to train and are high in accuracy. The formulation of such classifiers is still an active ongoing research topic in phoneme classification. One paradigm in formulating such classifiers attempts to increase the accuracies of the simpler classifiers with minimal sacrifice to their running times. The opposite paradigm attempts to increase the training speeds of the more complex classifiers with minimal sacrifice to their accuracies. The objective of this research is to develop a new centroid-based classifier that builds upon the simpler nearest centroid classifier by incorporating a new discriminative locally-adaptive training procedure developed from recent advances in machine learning. This new classifier, which is referred to as the discriminative locally-adaptive nearest centroid (DLANC) classifier, achieves much higher accuracies as compared to the nearest centroid classifier whilst having a relatively low computational complexity and being able to scale up to very large datasets.