<p>We investigate the use of noise degraded facial images in the application of facial expression recognition. In particular, we trained Gabor+SVMclassifiers to recognize facial expressions images with various types of noise. We applied Gaussian noise, Poisson noise, varying levels of salt and pepper noise, and speckle noise to noiseless facial images. Classifiers were trained with images without noise and then tested on the images with noise. Next, the classifiers were trained using images with noise, and then on tested both images that had noise, and images that were noiseless. Finally, classifiers were tested on images while increasing the levels of salt and pepper in the test set. Our results reflected distinct degradation of recognition accuracy. We also discovered that certain types of noise, particularly Gaussian and Poisson noise, boost recognition rates to levels greater than would be achieved by normal, noiseless images. We attribute this effect to the Gaussian envelope component of Gabor filters being sympathetic to Gaussian-like noise, which is similar in variance to that of the Gabor filters. Finally, using linear regression, we mapped a mathematical model to this degradation and used it to suggest how recognition rates would degrade further should more noise be added to the images.</p>

Road Networks play an important role in various applications such as urban and rural planning, infrastructure planning, transportation management, vehicle navigation. Extraction of Roads from Remote Sensed satellite images for updating road database in geographical information systems (GIS) is generally done manually by a human operator. However, manual extraction of roads is time consuming and labor intensive process. In the existing literature, there are a great number of researches published for the purpose of automating the road extraction process. However, automated processes still yield some erroneous and incomplete results and human intervention is still required. The aim of this research is to propose a framework for road network extraction from high spatial resolution multi-spectral imagery (MSI) to improve the accuracy of road extraction systems. The proposed framework begins with a spectral classification using One-class Support Vector Machines (SVM) and Gaussian Mixture Models (GMM) classifiers. Spectral Classification exploits the spectral signature of road surfaces to classify road pixels. Then, an iterative template matching filter is proposed to refine spectral classification results. K-medians clustering algorithm is employed to detect candidate road centerline points. Final road network formation is achieved by Markov Random Fields. The extracted road network is evaluated against a reference dataset using a set of quality metrics.

Multivariate pattern analysis (MVPA) of fMRI data has been growing in popularity due to its sensitivity to networks of brain activation. It is performed in a predictive modeling framework which is natural for implementing brain state prediction and real-time fMRI applications such as brain computer interfaces. Support vector machines (SVM) have been particularly popular for MVPA owing to their high prediction accuracy even with noisy datasets. Recent work has proposed the use of relevance vector machines (RVM) as an alternative to SVM. RVMs are particularly attractive in time sensitive applications such as real-time fMRI since they tend to perform classification faster than SVMs. Despite the use of both methods in fMRI research, little has been done to compare the performance of these two techniques. This study compares RVM to SVM in terms of time and accuracy to determine which is better suited to real-time applications.

Acoustic Event Detection / The detection of the Acoustic Events (AEs) naturally produced in a meeting room may help to describe the human and social activity. The automatic description of interactions between humans and environment can be useful for providing: implicit assistance to the people inside the room, context-aware and content-aware information requiring a minimum of human attention or interruptions, support for high-level analysis of the underlying acoustic scene, etc. On the other hand, the recent fast growth of available audio or audiovisual content strongly demands tools for analyzing, indexing, searching and retrieving the available documents. Given an audio document, the first processing step usually is audio segmentation (AS), i.e. the partitioning of the input audio stream into acoustically homogeneous regions which are labelled according to a predefined broad set of classes like speech, music, noise, etc. Acoustic event detection (AED) is the objective of this thesis work. A variety of features coming not only from audio but also from the video modality is proposed to deal with that detection problem in meeting-room and broadcast news domains. Two basic detection approaches are investigated in this work: a joint segmentation and classification using Hidden Markov Models (HMMs) with Gaussian Mixture Densities (GMMs), and a detection-by-classification approach using discriminative Support Vector Machines (SVMs). For the first case, a fast one-pass-training feature selection algorithm is developed in this thesis to select, for each AE class, the subset of multimodal features that shows the best detection rate. AED in meeting-room environments aims at processing the signals collected by distant microphones and video cameras in order to obtain the temporal sequence of (possibly overlapped) AEs that have been produced in the room. When applied to interactive seminars with a certain degree of spontaneity, the detection of acoustic events from only the audio modality alone shows a large amount of errors, which is mostly due to the temporal overlaps of sounds. This thesis includes several novelties regarding the task of multimodal AED. Firstly, the use of video features. Since in the video modality the acoustic sources do not overlap (except for occlusions), the proposed features improve AED in such rather spontaneous scenario recordings. Secondly, the inclusion of acoustic localization features, which, in combination with the usual spectro-temporal audio features, yield a further improvement in recognition rate. Thirdly, the comparison of feature-level and decision-level fusion strategies for the combination of audio and video modalities. In the later case, the system output scores are combined using two statistical approaches: weighted arithmetical mean and fuzzy integral. On the other hand, due to the scarcity of annotated multimodal data, and, in particular, of data with temporal sound overlaps, a new multimodal database with a rich variety of meeting-room AEs has been recorded and manually annotated, and it has been made publicly available for research purposes.

Current representation schemes for automatic text classification treat documents as syntactically unstructured collections of words (Bag-of-Words) or `concepts' (Bag-of-Concepts). Past attempts to encode syntactic structure have treated part-of-speech information as another word-like feature, but have been shown to be less effective than non-structural approaches. We propose a new representation scheme using Holographic Reduced Representations (HRRs) as a technique to encode both semantic and syntactic structure, though in very different ways. This method is unique in the literature in that it encodes the structure across all features of the document vector while preserving text semantics. Our method does not increase the dimensionality of the document vectors, allowing for efficient computation and storage. We present the results of various Support Vector Machine classification experiments that demonstrate the superiority of this method over Bag-of-Concepts representations and improvement over Bag-of-Words in certain classification contexts.

Current representation schemes for automatic text classification treat documents as syntactically unstructured collections of words (Bag-of-Words) or `concepts' (Bag-of-Concepts). Past attempts to encode syntactic structure have treated part-of-speech information as another word-like feature, but have been shown to be less effective than non-structural approaches. We propose a new representation scheme using Holographic Reduced Representations (HRRs) as a technique to encode both semantic and syntactic structure, though in very different ways. This method is unique in the literature in that it encodes the structure across all features of the document vector while preserving text semantics. Our method does not increase the dimensionality of the document vectors, allowing for efficient computation and storage. We present the results of various Support Vector Machine classification experiments that demonstrate the superiority of this method over Bag-of-Concepts representations and improvement over Bag-of-Words in certain classification contexts.

Protein engineering can be performed by combinatorial techniques (directed evolution) and data-driven methods using machine-learning algorithms. The main characteristic of directed evolution (DE) is the application of an effective and efficient screen or selection on a diverse mutant library. As it is important to have a diverse mutant library for the success of DE, we compared the performance of DNA-shuffling and recombination PCR on fluorescent proteins using sequence information as well as statistical methods. We found that the diversity of the libraries DNA-shuffling and recombination PCR generates were dependent on type of skew primers used and sensitive to nucleotide identity levels between genes. DNA-shuffling and recombination PCR produced libraries with different crossover tendencies, suggesting that the two protocols could be used in combination to produce better libraries. Data-driven protein engineering uses sequence, structure and function data along with analyzed empirical activity information to guide library design. Boolean Learning Support Vector Machines (BLSVM) to identify interacting residues in fluorescent proteins and the gene templates were modified to preserve interactions post recombination. By site-directed mutagenesis, recombination and expression experiments, we validated that BLSVM can be used to identify interacting residues and increase the fraction of active proteins in the library. As an extension to the above experiments, DE was applied on monomeric Red Fluorescent Proteins to improve its spectral characteristics and structure-guided protein engineering was performed on penicillin G acylase (PGA), an industrially relevant catalyst, to change its substrate specificity.

A subfield of artificial intelligence, machine learning (ML), is concerned with the development of algorithms that allow computers to &ldquo / learn&rdquo / . ML is the process of training a system with large number of examples, extracting rules and finding patterns in order to make predictions on new data points (examples). The most common machine learning schemes are supervised, semi-supervised, unsupervised and reinforcement learning. These schemes apply to natural language processing, search engines, medical diagnosis, bioinformatics, detecting credit fraud, stock market analysis, classification of DNA sequences, speech and hand writing recognition in computer vision, to encounter just a few. In this thesis, we focus on Support Vector Machines (SVMs) which is one of the most powerful methods currently in machine learning. As a first motivation, we develop a model selection tool induced into SVM in order to solve a particular problem of computational biology which is prediction of eukaryotic pro-peptide cleavage site applied on the real data collected from NCBI data bank. Based on our biological example, a generalized model selection method is employed as a generalization for all kinds of learning problems. In ML algorithms, one of the crucial issues is the representation of the data. Discrete geometric structures and, especially, linear separability of the data play an important role in ML. If the data is not linearly separable, a kernel function transforms the nonlinear data into a higher-dimensional space in which the nonlinear data are linearly separable. As the data become heterogeneous and large-scale, single kernel methods become insufficient to classify nonlinear data. Convex combinations of kernels were developed to classify this kind of data [8]. Nevertheless, selection of the finite combinations of kernels are limited up to a finite choice. In order to overcome this discrepancy, we propose a novel method of &ldquo / infinite&rdquo / kernel combinations for learning problems with the help of infinite and semi-infinite programming regarding all elements in kernel space. This will provide to study variations of combinations of kernels when considering heterogeneous data in real-world applications. Combination of kernels can be done, e.g., along a homotopy parameter or a more specific parameter. Looking at all infinitesimally fine convex combinations of the kernels from the infinite kernel set, the margin is maximized subject to an infinite number of constraints with a compact index set and an additional (Riemann-Stieltjes) integral constraint due to the combinations. After a parametrization in the space of probability measures, it becomes semi-infinite. We analyze the regularity conditions which satisfy the Reduction Ansatz and discuss the type of distribution functions within the structure of the constraints and our bilevel optimization problem. Finally, we adapted well known numerical methods of semiinfinite programming to our new kernel machine. We improved the discretization method for our specific model and proposed two new algorithms. We proved the convergence of the numerical methods and we analyzed the conditions and assumptions of these convergence theorems such as optimality and convergence.

In this study, a P300 based Brain-Computer Interface (BCI) system design is realized by the implementation of the Spelling Paradigm. The main challenge in these systems is to improve the speed of the prediction mechanisms by the application of different signal processing and pattern classification techniques in BCI problems. The thesis study includes the design and implementation of a 10 channel Electroencephalographic (EEG) data acquisition system to be practically used in BCI applications. The electrical measurements are realized with active electrodes for continuous EEG recording. The data is transferred via USB so that the device can be operated by any computer. v Wiener filtering is applied to P300 Speller as a signal enhancement tool for the first time in the literature. With this method, the optimum temporal frequency bands for user specific P300 responses are determined. The classification of the responses is performed by using Support Vector Machines (SVM&rsquo / s) and Bayesian decision. These methods are independently applied to the row-column intensification groups of P300 speller to observe the differences in human perception to these two visual stimulation types. It is observed from the investigated datasets that the prediction accuracies in these two groups are different for each subject even for optimum classification parameters. Furthermore, in these datasets, the classification accuracy was improved when the signals are preprocessed with Wiener filtering. With this method, the test characters are predicted with 100% accuracy in 4 trial repetitions in P300 Speller dataset of BCI Competition II. Besides, only 8 trials are needed to predict the target character with the designed BCI system.

Promoter classification is the task of separating promoter from non promoter sequences. Determining promoter regions where the transcription initiation takes place is important for several reasons such as improving genome annotation and defining transcription start sites. In this study, various promoter prediction methods called ProK-means, ProSVM, and 3S1C are proposed. In ProSVM and ProK-means algorithms, structural features of DNA sequences are used to distinguish promoters from non promoters. Obtained results are compared with ProSOM which is an existing promoter prediction method. It is shown that ProSVM is able to achieve greater recall rate compared to ProSOM results. Another promoter prediction methods proposed in this study is 3S1C. The difference of the proposed technique from existing methods is using signal, similarity, structure, and context features of DNA sequences in an integrated way and a hierarchical manner. In addition to current methods related to promoter classification, the similarity feature, which compares the promoter regions between human and other species, is added to the proposed system. We show that the similarity feature improves the accuracy. To classify core promoter regions, firstly, signal, similarity, structure, and context features are extracted and then, these features are classified separately by using Support Vector Machines. Finally, output predictions are combined using multilayer perceptron. The result of 3S1C algorithm is very promising.