Towards pose-robust face recognition on video

This thesis investigates face recognition in video under the presence of large pose variations. It proposes a solution that performs simultaneous detection of facial landmarks and head poses across large pose variations, employs discriminative modelling of feature distributions of faces with varying poses, and applies fusion of multiple classifiers to pose-mismatch recognition. Experiments on several benchmark datasets have demonstrated that improved performance is achieved using the proposed solution.

Score-level multibiometric fusion based on Dempster-Shafer theory incorporating uncertainty factors

While existing multi-biometic Dempster-Shafer the- ory fusion approaches have demonstrated promising perfor- mance, they do not model the uncertainty appropriately, sug- gesting that further improvement can be achieved. This research seeks to develop a unified framework for multimodal biometric fusion to take advantage of the uncertainty concept of Dempster- Shafer theory, improving the performance of multi-biometric authentication systems. Modeling uncertainty as a function of uncertainty factors affecting the recognition performance of the biometric systems helps to address the uncertainty of the data and the confidence of the fusion outcome. A weighted combination of quality measures and classifiers performance (Equal Error Rate) are proposed to encode the uncertainty concept to improve the fusion. We also found that quality measures contribute unequally to the recognition performance, thus selecting only significant factors and fusing them with a Dempster-Shafer approach to generate an overall quality score play an important role in the success of uncertainty modeling. The proposed approach achieved a competitive performance (approximate 1% EER) in comparison with other Dempster-Shafer based approaches and other conventional fusion approaches.

Injury narrative text classification : a preliminary study

Description of a patient's injuries is recorded in narrative text form by hospital emergency departments. For statistical reporting, this text data needs to be mapped to pre-defined codes. Existing research in this field uses the Naïve Bayes probabilistic method to build classifiers for mapping. In this paper, we focus on providing guidance on the selection of a classification method. We build a number of classifiers belonging to different classification families such as decision tree, probabilistic, neural networks, and instance-based, ensemble-based and kernel-based linear classifiers. An extensive pre-processing is carried out to ensure the quality of data and, in hence, the quality classification outcome. The records with a null entry in injury description are removed. The misspelling correction process is carried out by finding and replacing the misspelt word with a soundlike word. Meaningful phrases have been identified and kept, instead of removing the part of phrase as a stop word. The abbreviations appearing in many forms of entry are manually identified and only one form of abbreviations is used. Clustering is utilised to discriminate between non-frequent and frequent terms. This process reduced the number of text features dramatically from about 28,000 to 5000. The medical narrative text injury dataset, under consideration, is composed of many short documents. The data can be characterized as high-dimensional and sparse, i.e., few features are irrelevant but features are correlated with one another. Therefore, Matrix factorization techniques such as Singular Value Decomposition (SVD) and Non Negative Matrix Factorization (NNMF) have been used to map the processed feature space to a lower-dimensional feature space. Classifiers with these reduced feature space have been built. In experiments, a set of tests are conducted to reflect which classification method is best for the medical text classification. The Non Negative Matrix Factorization with Support Vector Machine method can achieve 93% precision which is higher than all the tested traditional classifiers. We also found that TF/IDF weighting which works well for long text classification is inferior to binary weighting in short document classification. Another finding is that the Top-n terms should be removed in consultation with medical experts, as it affects the classification performance.

A continental-scale tool for acoustic identification of European bats

Summary 1. Acoustic methods are used increasingly to survey and monitor bat populations. However, the use of acoustic methods at continental scales can be hampered by the lack of standardized and objective methods to identify all species recorded. This makes comparable continent-wide monitoring difficult, impeding progress towards developing biodiversity indicators, transboundary conservation programmes and monitoring species distribution changes. 2. Here we developed a continental-scale classifier for acoustic identification of bats, which can be used throughout Europe to ensure objective, consistent and comparable species identifications. We selected 1350 full-spectrum reference calls from a set of 15 858 calls of 34 European species, from EchoBank, a global echolocation call library. We assessed 24 call parameters to evaluate how well they distinguish between species and used the 12 most useful to train a hierarchy of ensembles of artificial neural networks to distinguish the echolocation calls of these bat species. 3. Calls are first classified to one of five call-type groups, with a median accuracy of 97·6%. The median species-level classification accuracy is 83·7%, providing robust classification for most European species, and an estimate of classification error for each species. 4. These classifiers were packaged into an online tool, iBatsID, which is freely available, enabling anyone to classify European calls in an objective and consistent way, allowing standardized acoustic identification across the continent. 5. Synthesis and applications. iBatsID is the first freely available and easily accessible continental- scale bat call classifier, providing the basis for standardized, continental acoustic bat monitoring in Europe. This method can provide key information to managers and conservation planners on distribution changes and changes in bat species activity through time.

Pollen image classification using the classifynder system: Algorithm comparison and a case study on New Zealand honey

We describe an investigation into how Massey University’s Pollen Classifynder can accelerate the understanding of pollen and its role in nature. The Classifynder is an imaging microscopy system that can locate, image and classify slide based pollen samples. Given the laboriousness of purely manual image acquisition and identification it is vital to exploit assistive technologies like the Classifynder to enable acquisition and analysis of pollen samples. It is also vital that we understand the strengths and limitations of automated systems so that they can be used (and improved) to compliment the strengths and weaknesses of human analysts to the greatest extent possible. This article reviews some of our experiences with the Classifynder system and our exploration of alternative classifier models to enhance both accuracy and interpretability. Our experiments in the pollen analysis problem domain have been based on samples from the Australian National University’s pollen reference collection (2,890 grains, 15 species) and images bundled with the Classifynder system (400 grains, 4 species). These samples have been represented using the Classifynder image feature set.We additionally work through a real world case study where we assess the ability of the system to determine the pollen make-up of samples of New Zealand honey. In addition to the Classifynder’s native neural network classifier, we have evaluated linear discriminant, support vector machine, decision tree and random forest classifiers on these data with encouraging results. Our hope is that our findings will help enhance the performance of future releases of the Classifynder and other systems for accelerating the acquisition and analysis of pollen samples.

A comparison of classification algorithms within the Classifynder pollen imaging system

We describe an investigation into how Massey University's Pollen Classifynder can accelerate the understanding of pollen and its role in nature. The Classifynder is an imaging microscopy system that can locate, image and classify slide based pollen samples. Given the laboriousness of purely manual image acquisition and identification it is vital to exploit assistive technologies like the Classifynder to enable acquisition and analysis of pollen samples. It is also vital that we understand the strengths and limitations of automated systems so that they can be used (and improved) to compliment the strengths and weaknesses of human analysts to the greatest extent possible. This article reviews some of our experiences with the Classifynder system and our exploration of alternative classifier models to enhance both accuracy and interpretability. Our experiments in the pollen analysis problem domain have been based on samples from the Australian National University's pollen reference collection (2890 grains, 15 species) and images bundled with the Classifynder system (400 grains, 4 species). These samples have been represented using the Classifynder image feature set. In addition to the Classifynder's native neural network classifier, we have evaluated linear discriminant, support vector machine, decision tree and random forest classifiers on these data with encouraging results. Our hope is that our findings will help enhance the performance of future releases of the Classifynder and other systems for accelerating the acquisition and analysis of pollen samples. © 2013 AIP Publishing LLC.

An evaluation of the neocognitron

We describe a sequence of experiments investigating the strengths and limitations of Fukushima's neocognitron as a handwritten digit classifier. Using the results of these experiments as a foundation, we propose and evaluate improvements to Fukushima's original network in an effort to obtain higher recognition performance. The neocognitron's performance is shown to be strongly dependent on the choice of selectivity parameters and we present two methods to adjust these variables. Performance of the network under the more effective of the two new selectivity adjustment techniques suggests that the network fails to exploit the features that distinguish different classes of input data. To avoid this shortcoming, the network's final layer cells were replaced by a nonlinear classifier (a multilayer perceptron) to create a hybrid architecture. Tests of Fukushima's original system and the novel systems proposed in this paper suggest that it may be difficult for the neocognitron to achieve the performance of existing digit classifiers due to its reliance upon the supervisor's choice of selectivity parameters and training data. These findings pertain to Fukushima's implementation of the system and should not be seen as diminishing the practical significance of the concept of hierarchical feature extraction embodied in the neocognitron. © 1997 IEEE.

Application of higher-order spectra for automated grading of diabetic maculopathy

Diabetic macular edema (DME) is one of the most common causes of visual loss among diabetes mellitus patients. Early detection and successive treatment may improve the visual acuity. DME is mainly graded into non-clinically significant macular edema (NCSME) and clinically significant macular edema according to the location of hard exudates in the macula region. DME can be identified by manual examination of fundus images. It is laborious and resource intensive. Hence, in this work, automated grading of DME is proposed using higher-order spectra (HOS) of Radon transform projections of the fundus images. We have used third-order cumulants and bispectrum magnitude, in this work, as features, and compared their performance. They can capture subtle changes in the fundus image. Spectral regression discriminant analysis (SRDA) reduces feature dimension, and minimum redundancy maximum relevance method is used to rank the significant SRDA components. Ranked features are fed to various supervised classifiers, viz. Naive Bayes, AdaBoost and support vector machine, to discriminate No DME, NCSME and clinically significant macular edema classes. The performance of our system is evaluated using the publicly available MESSIDOR dataset (300 images) and also verified with a local dataset (300 images). Our results show that HOS cumulants and bispectrum magnitude obtained an average accuracy of 95.56 and 94.39 % for MESSIDOR dataset and 95.93 and 93.33 % for local dataset, respectively.

Turning gaming EEG peripherals into trainable brain computer interfaces

Companies such as NeuroSky and Emotiv Systems are selling non-medical EEG devices for human computer interaction. These devices are significantly more affordable than their medical counterparts, and are mainly used to measure levels of engagement, focus, relaxation and stress. This information is sought after for marketing research and games. However, these EEG devices have the potential to enable users to interact with their surrounding environment using thoughts only, without activating any muscles. In this paper, we present preliminary results that demonstrate that despite reduced voltage and time sensitivity compared to medical-grade EEG systems, the quality of the signals of the Emotiv EPOC neuroheadset is sufficiently good in allowing discrimina tion between imaging events. We collected streams of EEG raw data and trained different types of classifiers to discriminate between three states (rest and two imaging events). We achieved a generalisation error of less than 2% for two types of non-linear classifiers.

Speeding up AdaBoost Classifier with Random Projection

The development of techniques for scaling up classifiers so that they can be applied to problems with large datasets of training examples is one of the objectives of data mining. Recently, AdaBoost has become popular among machine learning community thanks to its promising results across a variety of applications. However, training AdaBoost on large datasets is a major problem, especially when the dimensionality of the data is very high. This paper discusses the effect of high dimensionality on the training process of AdaBoost. Two preprocessing options to reduce dimensionality, namely the principal component analysis and random projection are briefly examined. Random projection subject to a probabilistic length preserving transformation is explored further as a computationally light preprocessing step. The experimental results obtained demonstrate the effectiveness of the proposed training process for handling high dimensional large datasets.

Word Sense Discovery and Disambiguation

The work is based on the assumption that words with similar syntactic usage have similar meaning, which was proposed by Zellig S. Harris (1954,1968). We study his assumption from two aspects: Firstly, different meanings (word senses) of a word should manifest themselves in different usages (contexts), and secondly, similar usages (contexts) should lead to similar meanings (word senses). If we start with the different meanings of a word, we should be able to find distinct contexts for the meanings in text corpora. We separate the meanings by grouping and labeling contexts in an unsupervised or weakly supervised manner (Publication 1, 2 and 3). We are confronted with the question of how best to represent contexts in order to induce effective classifiers of contexts, because differences in context are the only means we have to separate word senses. If we start with words in similar contexts, we should be able to discover similarities in meaning. We can do this monolingually or multilingually. In the monolingual material, we find synonyms and other related words in an unsupervised way (Publication 4). In the multilingual material, we ?nd translations by supervised learning of transliterations (Publication 5). In both the monolingual and multilingual case, we first discover words with similar contexts, i.e., synonym or translation lists. In the monolingual case we also aim at finding structure in the lists by discovering groups of similar words, e.g., synonym sets. In this introduction to the publications of the thesis, we consider the larger background issues of how meaning arises, how it is quantized into word senses, and how it is modeled. We also consider how to define, collect and represent contexts. We discuss how to evaluate the trained context classi?ers and discovered word sense classifications, and ?nally we present the word sense discovery and disambiguation methods of the publications. This work supports Harris' hypothesis by implementing three new methods modeled on his hypothesis. The methods have practical consequences for creating thesauruses and translation dictionaries, e.g., for information retrieval and machine translation purposes. Keywords: Word senses, Context, Evaluation, Word sense disambiguation, Word sense discovery.

Focused Prefetching: Performance Oriented Prefetching Based On Commit Stalls

Loads that miss in L1 or L2 caches and waiting for their data at the head of the ROB cause significant slow down in the form of commit stalls. We identify that most of these commit stalls are caused by a small set of loads, referred to as LIMCOS (Loads Incurring Majority of COmmit Stalls). We propose simple history-based classifiers that track commit stalls suffered by loads to help us identify this small set of loads. We study an application of these classifiers to prefetching. The classifiers are used to train the prefetcher to focus on the misses suffered by LIMCOS. This, referred to as focused prefetching, results in a 9.8% gain in IPC over naive GHB based delta correlation prefetcher along with a 20.3% reduction in memory traffic for a set of 17 memory-intensive SPEC2000 benchmarks. Another important impact of focused prefetching is a 61% improvement in the accuracy of prefetches. We demonstrate that the proposed classification criterion performs better than other existing criteria like criticality and delinquent loads. Also we show that the criterion of focusing on commit stalls is robust enough across cache levels and can be applied to any prefetcher without any modifications to the prefetcher.

Validation-Based Sparse Gaussian Process Classifier Design

Gaussian processes (GPs) are promising Bayesian methods for classification and regression problems. Design of a GP classifier and making predictions using it is, however, computationally demanding, especially when the training set size is large. Sparse GP classifiers are known to overcome this limitation. In this letter, we propose and study a validation-based method for sparse GP classifier design. The proposed method uses a negative log predictive (NLP) loss measure, which is easy to compute for GP models. We use this measure for both basis vector selection and hyperparameter adaptation. The experimental results on several real-world benchmark data sets show better orcomparable generalization performance over existing methods.

Learning Optimal Discriminant Functions through a Cooperative Game of Automata

The problem of learning correct decision rules to minimize the probability of misclassification is a long-standing problem of supervised learning in pattern recognition. The problem of learning such optimal discriminant functions is considered for the class of problems where the statistical properties of the pattern classes are completely unknown. The problem is posed as a game with common payoff played by a team of mutually cooperating learning automata. This essentially results in a probabilistic search through the space of classifiers. The approach is inherently capable of learning discriminant functions that are nonlinear in their parameters also. A learning algorithm is presented for the team and convergence is established. It is proved that the team can obtain the optimal classifier to an arbitrary approximation. Simulation results with a few examples are presented where the team learns the optimal classifier.

Automatic ICD-10 classification of cancers from free-text death certificates

Objective Death certificates provide an invaluable source for cancer mortality statistics; however, this value can only be realised if accurate, quantitative data can be extracted from certificates – an aim hampered by both the volume and variable nature of certificates written in natural language. This paper proposes an automatic classification system for identifying cancer related causes of death from death certificates. Methods Detailed features, including terms, n-grams and SNOMED CT concepts were extracted from a collection of 447,336 death certificates. These features were used to train Support Vector Machine classifiers (one classifier for each cancer type). The classifiers were deployed in a cascaded architecture: the first level identified the presence of cancer (i.e., binary cancer/nocancer) and the second level identified the type of cancer (according to the ICD-10 classification system). A held-out test set was used to evaluate the effectiveness of the classifiers according to precision, recall and F-measure. In addition, detailed feature analysis was performed to reveal the characteristics of a successful cancer classification model. Results The system was highly effective at identifying cancer as the underlying cause of death (F-measure 0.94). The system was also effective at determining the type of cancer for common cancers (F-measure 0.7). Rare cancers, for which there was little training data, were difficult to classify accurately (F-measure 0.12). Factors influencing performance were the amount of training data and certain ambiguous cancers (e.g., those in the stomach region). The feature analysis revealed a combination of features were important for cancer type classification, with SNOMED CT concept and oncology specific morphology features proving the most valuable. Conclusion The system proposed in this study provides automatic identification and characterisation of cancers from large collections of free-text death certificates. This allows organisations such as Cancer Registries to monitor and report on cancer mortality in a timely and accurate manner. In addition, the methods and findings are generally applicable beyond cancer classification and to other sources of medical text besides death certificates.