Petia Georgieva | University of Aveiro (original) (raw)

Papers by Petia Georgieva

2021 International Conference on INnovations in Intelligent SysTems and Applications (INISTA), Aug 25, 2021

Recently, deep neural networks achieved state-of-the-art results on the automated diagnosis of sk... more Recently, deep neural networks achieved state-of-the-art results on the automated diagnosis of skin lesions. Both the availability of bigger and better datasets as well as major advancements in Convolutional Neural Network methodologies represent some of the reasons behind these results. While the former is powered by initiatives like the International Skin Imaging Collaboration (ISIC), the latter is potentiated by developments in CNN architectures and the rise of transfer learning. This paper addresses open research questions related to the effectiveness of transfer learning methods in the context of multi-class skin lesion classification. The results indicate that, depending on the way pre-trained models are re-purposed, recent CNN architectures can bring significant performance boosts on the overall performance of deep learning classifiers. Experiments also highlight the importance of a good dataset to train these models, and how class balancing through data augmentation can help ease this requirement. Furthermore, experimentation with different models shows that ensembles can bring an edge over single-model approaches. Finally, this work presents a competitive single- and multi-model approach to the ISIC 2019 challenge.

Recent years have seen significant progress in the automatic diagnosis of pigmented skin lesions,... more Recent years have seen significant progress in the automatic diagnosis of pigmented skin lesions, including advances in self-surveillance technologies accessible to patients and computer-aided diagnosis (CAD) tools for dermatologists. Rapid advances in mobile technologies and applications are playing a central role in providing educational aids and self-surveillance tools for patient use. At the same time, machine learning, specifically, deep learning is a fast-growing field that is being used for multiple medical imaging related problems, such as skin lesions classification. Recent studies based on deep networks produced promising results which have the potential to change the landscape of skin lesion diagnosis. Systems created based on these new advancements aim to provide support for both dermatologists in the decision making process and for patients that do not have access to skin professionals. This paper focuses on the current state of automated skin lesion diagnosis, while also providing a comprehensive view into the challenges and opportunities in dermatology care.

2021 International Conference on INnovations in Intelligent SysTems and Applications (INISTA), Aug 25, 2021

Recent years have seen significant advances in automated diagnosis systems for medical imaging ta... more Recent years have seen significant advances in automated diagnosis systems for medical imaging tasks aimed to support the decision-making process. More specifically, Convolutional neural networks (CNN) show remarkable performance in tasks such as multi-class skin lesion classification using images. However, concerns remain about the deployment of such models, as real-world test data distribution can significantly differ from the distribution of the training data. In other words, models can classify unknown samples as known classes with high confidence, which could lead to catastrophic mistakes. In line with these concerns, this paper focuses on accessing the current methods to detect out-of-training distribution samples in the context of skin lesion classification. The results contribute towards the understanding of the effectiveness of out-of-distribution detection methods.

Intention inference from observation of human actions is an essential ability for robots performi... more Intention inference from observation of human actions is an essential ability for robots performing interactive tasks. This paper studies the role of early anticipation skills to improve the performance of a robotic system playing ball catching with a human partner. The source of anticipatory information results from the observation of the thrower's motion before the ball is released. For that purpose, a feed-forward neural network is trained to estimate the initial position and velocity of the ball in-flight given a sequence of observations during the throwing phase. The proposed approach outperforms up to 20% the classical methodology in which the generation of predictions solely relies upon the available information during the flight phase. Several simulation results demonstrate the added value of early anticipation skills from the viewpoint of ball catching performance.

Lecture Notes in Computer Science, 2015

In this paper a new adaptive Brain Computer Interface (BCI) architecture is proposed that allows ... more In this paper a new adaptive Brain Computer Interface (BCI) architecture is proposed that allows to autonomously adapt the BCI parameters in malfunctioning situations. Such situations are detected by discriminating EEG Error Potentials and when necessary the BCI mode is switched back to the training stage in order to improve its performance. First, the modules of the adaptive BCI are presented, then the scenarios for identification of the user reaction to intentionally introduced errors are discussed and finally promising preliminary results are commented. The proposed concept has the potential to increase the reliability of BCI systems.

Robotics, Oct 15, 2021

This article is an open access article distributed under the terms and conditions of the Creative... more This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY

Pattern Recognition and Image Analysis

2020 International Conference Automatics and Informatics (ICAI), 2020

This paper proposes a control system to enhance the performance of a solar panel. A two axes mech... more This paper proposes a control system to enhance the performance of a solar panel. A two axes mechanism is developed that tilts and turns the solar panel to face the highest intensity of light. The system was designed in LabVIEW and implemented on the Arduino Mega 2560. The physical model of the system was built using servo motors and photoresistors. The pilot plant was tested by applying a source of light from various directions and monitoring its response. The solar panel was able to face towards the highest intensity of light with high level of precision.

2018 IEEE International Conference on Autonomous Robot Systems and Competitions (ICARSC), 2018

2020 IEEE 16th International Conference on Control & Automation (ICCA), 2020

In this paper we propose an autopilot control strategy for a small radio controlled aircraft aimi... more In this paper we propose an autopilot control strategy for a small radio controlled aircraft aiming to follow a circular path with a desired radius. In contrast to the usual way to compute the error as the distance between the plane current space position and the desired flight trajectory path (for example a circle of radius r0), here we propose an alternative way to compute the scalar quantity E(t)) such that the aircraft moves in a spiral path of a decreasing/increasing with time radius to approach the circle with desired radius r0. Our particular choice of error term is intended to cause only gradual changes in the aircraft’s direction so as to prevent unstable flight due to over-control.

This paper describes a control system to enhance the performance of a solar panel. A two-axis mec... more This paper describes a control system to enhance the performance of a solar panel. A two-axis mechanism is developed that tilts and turns the solar panel to face the highest intensity of light. The system was designed in LabVIEW, and implemented on the Arduino Mega 2560. The physical model of the system was built using servo motors and photoresistors. The pilot plant was tested by applying a source of light from various directions and monitoring its response. The solar panel was able to face towards the highest intensity of light with high level of precision.

2016 IEEE International Black Sea Conference on Communications and Networking (BlackSeaCom), 2016

In this paper we present an automatic face recognition system based on incremental Singular Value... more In this paper we present an automatic face recognition system based on incremental Singular Values Decomposition (SVD) and subject dependent Hidden Markov Models (HMM). For each subject, an individual HMM is trained with features, extracted from the orthogonal decomposition (SVD) of the subject's training images. The main advantage of the proposed SVD-HMM recognition system is the robustness against image dimensionality reduction. The system was tested on two benchmark face datasets — the Olivetti Research Laboratory (ORL) and the YALE database. The SVD-HMM was further compared with a standard SVD face recognition. SVD applied to the original (full size) images performs similarly to the SVD-HMM applied to the compressed (half of the original size) images. SVD degrades rapidly when the image is compressed.

2021 International Conference on INnovations in Intelligent SysTems and Applications (INISTA), 2021

2016 IEEE Symposium Series on Computational Intelligence (SSCI), 2016

We consider the dynamic EEG source localization problem with additional constraints on the expect... more We consider the dynamic EEG source localization problem with additional constraints on the expected value of the state. In dynamic EEG source localization, the source brains, also called dipoles, are not stationary but vary over time. Moreover, given our specific EEG experiment, we expect the dipoles to be located within a certain area of the brain (here, the visual cortex). We formulate this constrained dynamic source localization problem as a constrained non-linear state-estimation problem. Particle filters (PFs) are nowadays the state-of-theart in optimal non-linear and non-Gaussian state estimation. However, PFs cannot handle additional constraints on the state that cannot be incorporated within the system model. In this case, the additional constraint is on the mean of the state, which means that realizations of the state, also called particles within the PF framework, may or may not satisfy the constraint. However, the state must satisfy the constraint on average. This is indeed the case when tracking brain dipoles from EEG experiments that try to target a specific cortex of the brain. Such constraints on the mean of the state are hard to deal with because they reflect global constraints on the posterior density of the state. The popular solution of constraining every particle in the PF may lead either to a stronger condition or to a different (unrelated) condition; both of which result in incorrect estimation of the state. We propose the Iterative Mean Density Truncation (IMeDeT) algorithm, which inductively samples particles that are guaranteed to satisfy the constraint on the mean. Application of IMeDeT on synthetic and real EEG data shows that incorporating a priori constraints on the state improves the tracking accuracy as well as the convergence rate of the tracker.

2015 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2015

Localization of the brain neural generators that create Electroencephalographs (EEGs) has been an... more Localization of the brain neural generators that create Electroencephalographs (EEGs) has been an important problem in clinical, research and technological applications related to the brain. The active regions in the brain are modeled as equivalent current dipoles, and the positions and moments of these dipoles or brain sources are estimated. So far, the brain dipoles are assumed to be fixed or time-invariant. However, recent neurological studies are showing that brain sources are not static but vary (in terms of location and moment) depending on various internal and external stimuli. This paper presents a shift in the current paradigm of brain source localization by considering dynamic sources in the brain. We formulate the brain source estimation problem from EEG measurements as a (nonlinear) state-space model. We use the Particle Filter (PF), essentially a sequential Monte Carlo method, to track the trajectory of the moving dipoles in the brain. We further address the "curse of dimensionality," issue of the PF by taking advantage of the structure of the EEG state-space model, and marginalizing out the linearly evolving states. A Kalman Filter is used to optimally estimate the linear elements, whereas the PF is used to track only the non-linear components. This technique reduces the dimension of the problem; thus exponentially reducing the computational cost. Our simulation results show that, where the PF fails, the Marginalized PF is able to successfully track two dipoles in the brain with only 500 particles.

2015 International Symposium on Innovations in Intelligent SysTems and Applications (INISTA), 2015

The objective of this study is to explore the effectiveness of three digital shopping platforms (... more The objective of this study is to explore the effectiveness of three digital shopping platforms (Plain Interactive, Marker-based Augmented Reality and Markerless Augmented Reality), on the impressions and purchase intentions of consumers. The study is mainly interested in analysing whether intelligent shopping platforms with AR elements provide any added advantage to an advertised product in the form of favourable attitude or a stronger purchase impulse. During the tests with the three shopping platforms, quantitative data was collected via computerised questionnaire. High and Low class users were statistically extracted, corresponding to the high or low probability to buy or recommend the advertised brand. The results show that Markerless AR system clearly outperforms the Marker-based AR and the Plain Interactive in terms of positive attitude from the users. The second better performing system is the Marker-based AR, which closely follows the Markerless AR, while the Plain Interactive system obtains least approval.

2014 International Joint Conference on Neural Networks (IJCNN), 2014

In this paper, we propose a statistical approach to reconstruct the brain neuronal activity based... more In this paper, we propose a statistical approach to reconstruct the brain neuronal activity based only on recorded EEG data. The brain zones with the strongest activity are expressed at a macro level by a few number of active brain dipoles. Normally, for solving the EEG inverse problem, fixed dipole locations are assumed, independently of the different stimuli that excite the brain. The proposed particle filter (PF) framework presents a shift in the current paradigm by estimating dynamic brain dipoles, which may vary from one location to another in the brain depending on internal/external stimuli that may affect the brain. Also, in contrast to previous solutions, the proposed PF algorithm estimates simultaneously, the number of the active dipoles, their moving locations and their respective oscillations in the three dimensional head geometry.