Feature Selection for Speech Emotion Recognition in Spanish and Basque: On the Use of Machine Learning to Improve Human-Computer Interaction

Study of emotions in human-computer interaction is a growing research area. This paper shows an attempt to select the most significant features for emotion recognition in spoken Basque and Spanish Languages using different methods for feature selection. RekEmozio database was used as the experimental data set. Several Machine Learning paradigms were used for the emotion classification task. Experiments were executed in three phases, using different sets of features as classification variables in each phase. Moreover, feature subset selection was applied at each phase in order to seek for the most relevant feature subset. The three phases approach was selected to check the validity of the proposed approach. Achieved results show that an instance-based learning algorithm using feature subset selection techniques based on evolutionary algorithms is the best Machine Learning paradigm in automatic emotion recognition, with all different feature sets, obtaining a mean of 80,05% emotion recognition rate in Basque and a 74,82% in Spanish. In order to check the goodness of the proposed process, a greedy searching approach (FSS-Forward) has been applied and a comparison between them is provided. Based on achieved results, a set of most relevant non-speaker dependent features is proposed for both languages and new perspectives are suggested.

On the Selection of Non-Invasive Methods Based on Speech Analysis Oriented to Automatic Alzheimer Disease Diagnosis

The work presented here is part of a larger study to identify novel technologies and biomarkers for early Alzheimer disease (AD) detection and it focuses on evaluating the suitability of a new approach for early AD diagnosis by non-invasive methods. The purpose is to examine in a pilot study the potential of applying intelligent algorithms to speech features obtained from suspected patients in order to contribute to the improvement of diagnosis of AD and its degree of severity. In this sense, Artificial Neural Networks (ANN) have been used for the automatic classification of the two classes (AD and control subjects). Two human issues have been analyzed for feature selection: Spontaneous Speech and Emotional Response. Not only linear features but also non-linear ones, such as Fractal Dimension, have been explored. The approach is non invasive, low cost and without any side effects. Obtained experimental results were very satisfactory and promising for early diagnosis and classification of AD patients.

Recognition of Membrane-Bound Fusion-Peptide/MPER Complexes by the HIV-1 Neutralizing 2F5 Antibody : Implications for Anti-2F5 Immunogenicity

13 p.

Study of an object recognition algorithm

This project introduces an improvement of the vision capacity of the robot Robotino operating under ROS platform. A method for recognizing object class using binary features has been developed. The proposed method performs a binary classification of the descriptors of each training image to characterize the appearance of the object class. It presents the use of the binary descriptor based on the difference of gray intensity of the pixels in the image. It shows that binary features are suitable to represent object class in spite of the low resolution and the weak information concerning details of the object in the image. It also introduces the use of a boosting method (Adaboost) of feature selection al- lowing to eliminate redundancies and noise in order to improve the performance of the classifier. Finally, a kernel classifier SVM (Support Vector Machine) is trained with the available database and applied for predictions on new images. One possible future work is to establish a visual servo-control that is to say the reac- tion of the robot to the detection of the object.

Detección automática de presencia/ausencia de atún en imágenes obtenidas mediante sonar de largo alcance a bordo de buques pesqueros y aplicación de Optical Character Recognition (OCR) para la extracción de parámetros

En esta tesis de máster se presenta una metodología para el análisis automatizado de las señales del sonar de largo alcance y una aplicación basada en la técnica de reconocimiento óptico de Optical Character Recognition, caracteres (OCR). La primera contribución consiste en el análisis de imágenes de sonar mediante técnicas de procesamiento de imágenes. En este proceso, para cada imagen de sonar se extraen y se analizan las regiones medibles, obteniendo para cada región un conjunto de características. Con la ayuda de los expertos, cada región es identi cada en una clase (atún o no-atún). De este modo, mediante el aprendizaje supervisado se genera la base de datos y, a su vez, se obtiene un modelo de clasi cación. La segunda contribución es una aplicación OCR que reconoce y extrae de las capturas de pantalla de imágenes de sonar, los caracteres alfanuméricos correspondientes a los parámetros de situación (velocidad, rumbo, localización GPS) y la confi guración de sonar (ganancias, inclinación, ancho del haz). El objetivo de este proceso es el de maximizar la e ficiencia en la detección de atún en el Golfo de Vizcaya y dar el primer paso hacia el desarrollo de un índice de abundancia de esta especie, el cual esté basado en el procesamiento automático de las imágenes de sonar grabadas a bordo de la ota pesquera durante su actividad pesquera rutinaria.

A detailed investigation of classification methods for vowel speech imagery recognition

Accurate and fast decoding of speech imagery from electroencephalographic (EEG) data could serve as a basis for a new generation of brain computer interfaces (BCIs), more portable and easier to use. However, decoding of speech imagery from EEG is a hard problem due to many factors. In this paper we focus on the analysis of the classification step of speech imagery decoding for a three-class vowel speech imagery recognition problem. We empirically show that different classification subtasks may require different classifiers for accurately decoding and obtain a classification accuracy that improves the best results previously published. We further investigate the relationship between the classifiers and different sets of features selected by the common spatial patterns method. Our results indicate that further improvement on BCIs based on speech imagery could be achieved by carefully selecting an appropriate combination of classifiers for the subtasks involved.

Skeleton based visual pattern recognition. Applications to tabletop interaction

211 p. :il.

Application of Entropy and Fractal Dimension Analyses to the Pattern Recognition of Contaminated Fish Responses in Aquaculture

The objective of the work was to develop a non-invasive methodology for image acquisition, processing and nonlinear trajectory analysis of the collective fish response to a stochastic event. Object detection and motion estimation were performed by an optical flow algorithm in order to detect moving fish and simultaneously eliminate background, noise and artifacts. The Entropy and the Fractal Dimension (FD) of the trajectory followed by the centroids of the groups of fish were calculated using Shannon and permutation Entropy and the Katz, Higuchi and Katz-Castiglioni's FD algorithms respectively. The methodology was tested on three case groups of European sea bass (Dicentrarchus labrax), two of which were similar (C1 control and C2 tagged fish) and very different from the third (C3, tagged fish submerged in methylmercury contaminated water). The results indicate that Shannon entropy and Katz-Castiglioni were the most sensitive algorithms and proved to be promising tools for the non-invasive identification and quantification of differences in fish responses. In conclusion, we believe that this methodology has the potential to be embedded in online/real time architecture for contaminant monitoring programs in the aquaculture industry.

Frame-level features conveying phonetic information for language and speaker recognition

150 p.

Leukemia-Associated Mutations in Nucleophosmin Alter Recognition by CRM1: Molecular Basis of Aberrant Transport

Nucleophosmin (NPM) is a nucleocytoplasmic shuttling protein, normally enriched in nucleoli, that performs several activities related to cell growth. NPM mutations are characteristic of a subtype of acute myeloid leukemia (AML), where mutant NPM seems to play an oncogenic role. AML-associated NPM mutants exhibit altered subcellular traffic, being aberrantly located in the cytoplasm of leukoblasts. Exacerbated export of AML variants of NPM is mediated by the nuclear export receptor CRM1, and due, in part, to a mutationally acquired novel nuclear export signal (NES). To gain insight on the molecular basis of NPM transport in physiological and pathological conditions, we have evaluated the export efficiency of NPM in cells, and present new data indicating that, in normal conditions, wild type NPM is weakly exported by CRM1. On the other hand, we have found that AML-associated NPM mutants efficiently form complexes with CRM1HA (a mutant CRM1 with higher affinity for NESs), and we have quantitatively analyzed CRM1HA interaction with the NES motifs of these mutants, using fluorescence anisotropy and isothermal titration calorimetry. We have observed that the affinity of CRM1HA for these NESs is similar, which may help to explain the transport properties of the mutants. We also describe NPM recognition by the import machinery. Our combined cellular and biophysical studies shed further light on the determinants of NPM traffic, and how it is dramatically altered by AML-related mutations.

Dealing with the Effects of Sensor Displacement in Wearable Activity Recognition

Most wearable activity recognition systems assume a predefined sensor deployment that remains unchanged during runtime. However, this assumption does not reflect real-life conditions. During the normal use of such systems, users may place the sensors in a position different from the predefined sensor placement. Also, sensors may move from their original location to a different one, due to a loose attachment. Activity recognition systems trained on activity patterns characteristic of a given sensor deployment may likely fail due to sensor displacements. In this work, we innovatively explore the effects of sensor displacement induced by both the intentional misplacement of sensors and self-placement by the user. The effects of sensor displacement are analyzed for standard activity recognition techniques, as well as for an alternate robust sensor fusion method proposed in a previous work. While classical recognition models show little tolerance to sensor displacement, the proposed method is proven to have notable capabilities to assimilate the changes introduced in the sensor position due to self-placement and provides considerable improvements for large misplacements.