sinc(i) Research Center for Signals, Systems and Computational Intelligence

Respiratory signal processing

We are developing advanced techniques of signal processing and pattern recognition for automatic detection and classification of sleep breathing disorders. These respiratory disorders produce a fragmentation of sleep, causing daytime sleepiness, a reduction of reaction times, lack of concentration, and cardiac problems, among others. The prevalence in the population is very high, but as the studies to detect it are expensive and require of a sleep laboratory where the patient has to sleep all night, the disease is undiagnosed. The research project aims at producing techniques of rapid screening for these breathing anomalies, producing less invasive and easy to use devices, to make the sleep disorders study popular.

Blind Source Separation

This technique aims at obtaining estimation of the sources that produce some measurable field, given a set of distributed measurements of the field itself. Specifically applied to audio sources, the field is a sound field measured through several microphones located remotely in some specified locations. The problem is known as the "cocktail party problem", as in a party usually there are a lot of sound sources and yet human beings are able to focus their attention and listen to some specific source, like the person to which one is having a conversation. This problem is very difficult to solve, as in the general case, the sound field is affected by the room characteristics (the reverberation phenomenon), the sources and sensors can be arbitrarily located, and in the most general case, the sources and/or sensors can be moving in the space. We have developed several techniques for blind source separation of audio sources, two of which have been presented as patent applications in Japan. There are multiple applications for such techniques, for example in teleconference systems, human-machine interface devices, remote controlling of home appliances, among others.

ECG signal processing

The group was involved in the development of a library for digital signal processing and pattern recognition routines for automatic processing and classification of Holter ECG recordings.

The routines included several basic processing like different types of filters, baseline removal, noisy areas detection, and also more advanced techniques like morphological classification, QRS detection, ST deviation measurement, heart rate variability analysis in time and frequency domain, among others. The product is nowadays sold in Latin America and Europe. For more information the developed device can be found at: http://www.eccosur.com/holter.php

Emotion Recognition

Currently, the constant evolution of information technologies makes more necessary the human-machine interaction improvement. Human-machine interaction systems based on speech are capable of recognizing "what was said" and "who said this", using speech recognition and speaker identification techniques. However, words only contribute with 7% of information in human communications whereas paralinguistic information (pitch, volume, emotions, etc.) contributes with the 40%. It could be known "what was the emotional state when she/he said" just by adding an emotion recognition system to the general interaction system.

The recognition of emotions is a multi-disciplinary research area that has received great interest over the last years. Automatic recognition of speaker emotional state aims to achieve a more natural interaction between humans and machines. Also, it could be used to make the computer act according to the actual human emotion.

This could be applied to: security applications using detection of the emotional manifestation of fear in abnormal situations; systems for real-life emotion detection using a corpus of agent-client spoken dialogs from a medical emergency call center; a framework to support automatic diagnosis of psychiatric diseases; among others.

Speech Recognition

The process of recognition converts an acoustic signal, captured by a microphone, into a set of words or transcriptions. The challenge in automatic speech recognition is how to make many different information sources (acoustics, phonetics, phonology, lexical, syntactic, semantic and pragmatic) to cooperate in order to obtain an interpretation of the message.

Speech recognition applications include voice dialing, domestic appliance control, content-based spoken audio search, simple data entry (e.g., entering a credit card number) and speech-to-text processing (e.g., word processing).

Speaker Recognition

Automatic speaker recognition is the computer task of identifying a person from a spoken phrase. Speaker verification is the use of the machine to accept or reject a speaker's claimed identity, while in speaker identification the system decides the speaker's identity from a finite set of persons, with no identity claim. In both, verification and identification, the speaker's utterance is first analyzed, to extract some characteristic features, which are then compared with pre-trained speaker models. A speaker recognition system is text-independent if the user can enunciate an arbitrary sentence, and it is text-dependent if the user enunciates always the same sentence. Applications of speaker recognition include access control (e.g., to databases, restricted areas), information services and telephone banking.

Bioacustics and agriculture

Accurate measurements of feeding behavior are essential for a reliable management and investigation of grazing ruminants. An indication of animal health and welfare can be obtained by monitoring grazing and rumination activities, because ruminants have a daily chewing requirement to maintain a healthy rumen environment. In this application we propose novel methods to analyze and automatically recognize sound signals of chewing and biting in cows and sheeps, including pasture species identification and dry matter intake estimation. For the automatic segmentation and classification of acoustical ingestive behaviour these methods use appropriate acoustic representations and ad-hoc statistical modelling. Acoustic monitoring provides the most accurate quantification of chewing, and could be developed into a routine method to monitor animals such as dairy cows that are subject to the stresses of extremely high productivity.

Bioinformatics

Data mining on tomato fruits

The volume of information derived from postgenomic technologies is rapidly increasing. Due to the amount of data involved, novel computational methods are needed for the analysis and knowledge discovery into the massive data sets produced by these new technologies. We are working in biological data integration and discovery of a-priori unknown relationships between gene expression and metabolite variations, involving transcriptomic and metabolomic profiles from introgression lines of tomato fruits.

Download the *omeSOM software, a tool designed for data mining of metabolic and transcriptional datasets

Chromosome Recognition

A basic cytogenetic task is the karyotyping of a cell, where the chromosomes are identified and labelled for clinical analysis. It constitutes a very laborious practice, so in the last years different approaches to an automatic system were proposed by means of pattern recognition techniques.

The classifiers we developed are specialized in capture the variabilities in the dark and light band patterns of the stained chromosomes. For this purpose, different measures are introduced based on the digital image processing of the microphotograpies.

The classification is carried out by means of connexionist models (multilayer Perceptron and Elman networks) and continuous hidden Markov models, adapted to achieve a frame-by-frame analysis along the chromosome.

Also, an algorithm for contextual classification is proposed in order to emulate the performance of the expert in the clinical routine, who classify all the chromosomes of a cell as a whole taking into account the expected number of chromosomes in each karyotype group.

Face Recognition

Over the last years, face recognition has become one of the most popular biometric technologies. Its objective is to automatically identify a person from a picture or a frame in a video source.

A complete automatic face recognition system includes the following stages: face detection, face representation and face classification.

Face detection refers to finding a human face in the scene by means of image processing techniques. Here, we developed a face tracker based on color segmentation and eye detection. The output of this module is a face normalized both in size and brigthness.

The face is then represented through an appropiate feature extraction method, which provides useful information for the classification. This task is carried out by means of the classic eigenfaces technique.

The final stage corresponds to the recognition itself using some classifier, designed according to the previous extracted data. We developed an array of multilayer perceptron neural networks trained with a novel no-class resampling strategy. This method takes into account the balance problem between class and no-class examples and increase the generalization capabilities.

Automatic hand sign recognition

A real-time method to identify hand signs was developed. The user does signs for a standard web-cam, with natural illumination and without a special equipment. Some applications for this system include machine automatic control, interactive software to teach language of hand signs, interactive games, etc.

Photovoltaic System Optimization

An important growth in the power of the photovoltaic systems connected to a grid has recently been observed. In spite of the advances in module technology, the problems in the system design increased, especially regarding the surface of the earth they occupy. We developed a complete model for plant simulation with different wiring diagrams and under dynamic shading. Results obtained from simulations showed that the configuration with the lowest performance was that of only one serial-parallel group, whereas the highest efficiency corresponded to a design of groups of modules in parallel connected then in series. The simulation model proposed allows exploring different alternatives of wiring modules and finding the most efficient configurations for photovoltaic plants of medium and high power.

The main objective in this application is to develop intelligent algorithms that optimize the performance of photovoltaic centrals of medium and high power, in real time and considering both the load variations as the climate changes that affect the performance of individual panels, finding optimum schemes of plant design from the point of view of energy efficiency.

Development of an unmanned and autonomous hovercraft

The aim of this project is the development of an unmanned hovercraft with an autonomous navigation system and remote operation capabilities. The autonomous navigational system should be able of operating in non structured and dynamical environments in order to accomplish the programmed tasks. The vehicle should be able to perform complex and/or repetitive tasks in dangerous environments (hostile weather conditions, difficult navigation conditions, polluted environments conditions, among others) with minimal human operator actions. However, the operator can remotely control the vehicle or modify the programmed tasks sequence when it is required. In this way, the operator can carry out the task without being exposed to the environmental conditions, avoiding the risk. A fuzzy navigation system in an unknown and dynamic enviroment is showed in the video.