Advanced Techniques of Radar Detection in Non-Gaussian Background
Maria Greco, Professor, University of Pisa, IEEE Signal Processing Society Distinguished Lecturer

Abstract: For several decades, the Gaussian assumption on the disturbance modeling in radar systems has been widely used to deal with detection problems. But, in modern high-resolution radar systems, the disturbance cannot be modelled as Gaussian distributed and the classical detectors suffer from high losses.
In this talk, after a brief description of modern statistical and spectral models for high-resolution clutter, coherent optimum and sub-optimum detectors, designed for such a background, will be presented and their performance analyzed against a non-Gaussian disturbance. Different interpretations of the various detectors are provided that highlight the relationships and the differences among them.
After this first part, some discussion will be dedicated to how to make adaptive the detectors, by incorporating a proper estimate of the disturbance covariance matrix. Recent works on Maximum Likelihood and robust covariance matrix estimation have proposed different approaches such as the Approximate ML (or Fixed-Point) Estimator or the M-estimators. These techniques allow to improve the detection performance in terms of false alarm regulation and detection gain in SNR.
A plethora of results with simulated and real recorded data will be shown.

Signal Processing and Dynamical Systems on Weighted Lattices
Petros Maragos, Professor, National Technical University of Athens, Greece

Abstract: In this talk we will present a new unifying theoretical framework of nonlinear signal processing operators and dynamical systems that obey a superposition of a weighted max-min type and evolve on nonlinear spaces which we call complete weighted lattices. Special cases of such systems have found applications in morphological image analysis and vision scale-spaces, in control of discrete-event dynamical systems with minimax algebra, in speech recognition as weighted finite-state transducers, and in belief propagation on graphical models. Our theoretical approach establishes their representation in state and input-output spaces using monotone lattice operators, finds analytically their state and output responses using nonlinear convolutions of a weighted max-min type, studies their stability, and provides optimal solutions to solving max-min matrix equations. The talk will emphasize the main concepts and theoretical results in this broad field using weighted lattice algebra and will outline some application areas.
More information, related papers and current results can be found in http://cvsp.cs.ntua.gr .

Adaptive Sinusoidal Models and their applications in speech processing
Yannis Stylianou, Professor, University of Crete

Abstract:  Standard sinusoidal models have been shown to be successful in many applications in audio and speech processing. In this talk the passage from the standard non-adaptive and stationary sinusoidal models to the adaptive and non-stationary sinusoidal models will be presented in the context of speech analysis. Novel algorithms for the adaptive speech analysis will be presented and how they are related to the well known sinusoidal representation as well as to non-linear frequency estimators like the Newton-Gauss algorithm. Specific applications of the adaptive sinusoidal models in voice quality assessment will be presented as well their potential for well known speech technologies such as Automatic Speech Recognition, Text to Speech, will be discussed.

Human action recognition: Methods based on Gaussian Mixture Models and Conditional Random Fields
Christophoros Nikou, Associate Professor, University of Ioannina, Greece

Abstract: Human action and activity recognition play an important role in many applications such as video surveillance and human-computer interaction. At first, a brief introduction to the state of the art and an action recognition method relying on Gaussian mixtures of visual cues and a non-metric similarity function will be presented. Then, we will focus on modeling the behavior of a subject with a conditional random field (CRF) whose unary terms employ spatio-temporal features and the pairwise terms are based on kinematic features. This model is then extended, with the insertion of a hidden layer, to a hidden conditional random field (HCRF) and it is applied to multimodal (visual and audio) features. Moreover, as human movements are highly correlated with sound emissions, canonical correlation analysis is applied to estimate the temporal synchronization offset between the audio and video streams prior to fusion. Experiments performed on two human behavior recognition data sets including political speeches and human interactions from TV shows underpin the advantages of the proposed method compared with several baseline approaches.

Novel Techniques to Advance the Frontiers of Old Greek Document Recognition
Basilis Gatos, Research Director, National Center for Scientific Research “Demokritos”, Greece

Abstract: After many years of scholar study, old Greek machine-printed and handwritten collections continue to be an important source of novel information for scholars, concerning both the history of earlier times as well as the development of cultural documentation over the centuries. Although the accurate recognition of Latin machine-printed text is now considered largely a solved problem, recognition of old Greek documents is still the subject of active research. In this presentation, we will focus on our recent achievements for the recognition of Early Christian Greek manuscripts and old Greek polytonic (multi accent) machine-printed documents.  All these documents originate from St. Catherine’s Mount Sinai Monastery as well as from several Greek archives and libraries and contain a vast amount of valuable information. A robust indexing of these documents is essential for quick and efficient content exploitation of the valuable historical collections. The continuity in writing for characters of the same or consecutive words as well as the unique characteristics of the lower case script in Early Greek manuscripts guided us to develop a segmentation-free recognition technique as a fundamental assistance to old Greek handwritten manuscript OCR. Recognition of machine-printed document images having Greek polytonic characters is also a challenging task due the large number of existing character classes (more than 270). To this end, we have focused our research efforts towards an OCR engine for Greek polytonic documents that combines suitable pre-processing, segmentation, feature extraction and classification modules in order to support and facilitate current and future efforts in old Greek document digitization and processing.

Sparsity-aware adaptive parameter estimation: a Bayesian perspective
Athanasios Rontogiannis, Senior Researcher, National Observatory of Athens, Greece

Abstract: Adaptive estimation of time-varying signals and systems is a research field that has attracted significant attention in the statistical signal processing literature and has had a great impact in a plethora of applications. A large number of adaptive estimation techniques have been developed and analyzed during the past decades, which have the ability to process streaming data and provide real-time estimates of the parameters of interest in an online fashion. Recently, the interest in the area has been revived, thanks to the advancements in the compressive sensing (CS) field, which provide the means to exploit parameter sparsity in a time-varying environment.  By leveraging parameter sparsity, which naturally appears in most signals and systems, significant improvements in convergence rate and estimation performance of adaptive techniques can be achieved.
Unlike most state-of-the-art sparsity-aware adaptive estimation techniques, which are deterministic, in this presentation, we will provide an alternative approach stemming from a Bayesian perspective. In this context, both conventional and structured (group-) sparsity schemes will be explored.  In all these schemes, appropriate hierarchical Bayesian models are first defined where sparsity is imposed by assigning heavy-tailed sparsity-promoting priors to the parameters of interest. Then, an approximate Bayesian inference method, termed variational Bayes, is applied to obtain estimates of all parameters involved in the models. The resulting fully automated variational Bayes schemes will be first presented in a batch iterative form. Then, it will be shown that by properly exploiting the structure of the batch estimation task, new sparse online Bayesian algorithms can be derived. The most important feature of the proposed algorithms is that they completely eliminate the need for computationally costly parameter fine-tuning, a necessary ingredient of sparse adaptive deterministic algorithms. We will present extensive simulation results on adaptive channel estimation, which demonstrate the effectiveness of the new sparse adaptive variational Bayes algorithms against state-of-the-art deterministic techniques.

Cognitive Multimodal Processing: From Signal to Behavior
Alexandros Potamianos,  Associate Professor, National Technical University of Athens, Greece

Abstract: Affective computing, social and behavioral signal processing are emerging research disciplines that attempt to automatically label the emotional, social and cognitive state of humans using features extracted from audio-visual streams. I argue that this monumental task cannot succeed unless the particularities of the human cognitive processing are incorporated into our models, especially given that often the quantities we are called to model are either biased cognitive abstractions of the real world or altogether fictional creations of our cognition. A variety of cognitive processes that make computational modeling especially challenging are outlined, notably: 1) (joint) attention and saliency, 2) common ground, conceptual semantic spaces and representation learning, 3) fusion across time, modalities and cognitive representation layers, and 4) dual-system processing (system one vs. system two) and cognitive decision non-linearities. In this presentation, grand challenges are outlined and examples are given illustrating how to design models that are both high-performing and respect basic cognitive organization principles. It is shown that such models can achieve good generalization and representation power, as well as model cognitive biases, a prerequisite for modeling and predicting human behavior.