AREA 2. - ARTIFICIAL INTELLIGENCE AND COGNITIVE SCIENCE
Activity 2.1 – Computer vision
E. Catanzariti, R. Prevete, M. Santoro(graduate student
We started a new joint collaboration with INFN (MAGIC-5 Experiment). In this context the
development of an automated system to be used as support for the process of screening in the
diagnosis of lung cancer with CT images is under way. We are using AI and algorithmic
techniques in order to increase the sensitivity of detection in lung images obtained with a low
intensity spiral CT possible areas containing neoplasic lesions.
The vision system JVista has been improved, both in terms of the number of available functions
and of the efficiency. In this context we have implemented a number of clustering algorithms:
k-means, fuzzy c-means, spatially guided fuzzy c-means, bias corrected fuzzy c-means.
We have continued the work on image representation based on the Gabor transform.
A prototype CAD system able to detect massive lesions into CALMA and MIAS databases has
been implemented. With such a system we have validated the proposed segmentation schema,
prior to its integration within GPCALMA.
Program for 2005
We plan to test the segmentation schema of massive lesions with the remaining parts of
CALMA and MIAS databases.
We plan to address the problem of analysing and implementing image processing algorithms
based on probabilistic-generative models. In particular: an EM algorithm for the maximum
likelihood estimate of the density of mixtures of gaussians, a supervised algorithm for the
estimate of the density based on the histogram of class representatives, statistical classifiers as
“Maximum a posteriori”, “Standard likelood ratio”, and “Plug-in”.
We also intend to continue the work on the automatic interpretation of mammographic images
with the intent to develop a fuzzy classifier for the recognition and classification of stellate
lesions.
Activity 2.2 – Digital signal processing for multimedia
S. Cavaliere, G. Evangelista
The research has been developed along three main directions. First direction is the accurate
study of the sound production for a large class of musical instruments, in which dispersive
propagation takes place, producing acoustically significant deviation from the simple harmonic
structure in the sound spectrum. A transform method approach is proposed to solve the classic
fourth-order equations of stiff systems in order to reduce it to two second-order equations. By
introducing scattering boundary matrices, the eigenfrequencies are determined and their
dependency is discussed for the clamped, hinged, and intermediate cases.
The resulting pseudo-periodic feature of this class of sounds, including strings but also bars and
percussions, calls for complex algorithms for sound synthesis, which should be suitable for real
time implementation, and embody the feature of frequency dependent speed of propagation; the
proposed solution is found in the structure of digital waveguides in which the insertion of an
all-pass delay line generalizes a classical algorithm for the synthesis of ideally flexible vibrating
strings. Knowing the physical parameters, the synthesis can proceed using the generalized
structure. Another point of view is offered by Laguerre expansions and frequency warping,
which are introduced in order to show that a stiff system can be treated as a non-stiff one,
provided that the solutions are warped.
A second direction of research has focused on the problem of analysis and classification of
sound and audio signals; this problem, very relevant for different research fields from audio
browsing in Multimedia Archives to Computerized Audio Scene Analysis, has been
approached under the particular view point of automatic instrument identification. In this
direction intensive work has been done for the definition and extraction of relevant sound
features; these features could be divided into features regarding the whole evolution of the
sound and the short term features; focusing on the latter these are efficiently described by means
of the evolution of MFCC (mel frequency cepstral coefficients) over time. Extensive
simulations have been carried on for the classification, on the ground of these short time
signature, using data mining techniques, ranging from Kohonen maps to PCA (principal
component analysis) and comparing the results, in order to asses the role of these parameters in
the process of identification.
The third direction has focused the problem of separation of musical sound sources; this
problem has been solved at a certain degree even in the very complex case of overlapping
partials which is usual for sounds from musical instruments. The method developed in this
research makes use of the directional information, analysed in the form of phase delay,
integrated by means of the use of HRTFs (Head Related Transfer Functions) and is based
mainly on the detection of energy evolution of the isolated partials; the detected temporal
envelopes may then be grouped on the ground of the similarity of their time evolution, thus
providing a robust model for the original sound sources in the mixture.
G. Evangelista served as referee or member of scientific boards of the IEEE Transaction on
Signal Processing, Signal Processing Letters, Signal Processing (EURASIP), Journal of Applied
Signal Processing (EURASIP), DAFX Digital Audio Effects (European COST G6), IEEE
Multimedia & Expo, Speech Communications (Elsevier).
Program for 2005
In the field of sound and signal transformation, mainly frequency warping, we will work on the
development of a brand new range of frequency transformations, mapping the frequency axis in
a completely arbitrary way. This instrument besides the interest in the field of abstract unitary
signal transformations would have a practical application in the modification of the frequency
content of a signal, and should give rise to an algorithm which should allow real time. As shown
in the past frequency transforms are non causal in principle; however a suitable approximation
should be searched for, which circumventing the non causal nature of the transforms, allows a
real time realization, probably with an allowed fixed time delay. This will be valuable for the
real time synthesis of a large class of instrument sounds
As far as regards sound classification, in order to obtain a substantial improvement we will
study the relationship of the parameters with the pitch of the analysed sound; in fact a
relationship is readily demonstrated, and is already reported in the literature. Normalizing the
parameters against this dependency law will improve the classification task and make it more
robust against noise.
Regarding the third field of research , sound source sound identification, we will set up an
experimental apparatus in order to detect source sound direction in a robotic environment. The
algorithm already developed with respect to synthetic signals or anechoic environment will be
confronted against a real sound environment, thus analyzing the robustness against ambient
noise, echo and reverberation.
Activity 2.3 Hybrid systems, cybernetics and robotics
E. Burattini, E. Catanzariti, P. Coraggio(graduate student), G.Tamburini, R. Prevete, G. Trautteur
Within the project developed for the “Centro di Competenza sui Beni Culturali e Ambientali”
(Center for artistic and environmental resources) we realized on an Oracle platform a system for
the handling and management of texts, images, film footage and spatial data on dgw files
generated by Autocad.
On the robotics line we worked on autonomous robotic systems capable of reactive behaviour in
real environments. We have been exploring the possible unification of the cognitive approaches
“active perception” and “expected perception”. At the moment , in cooperation with the Robotic
Group of the Faculty of Engineering of our Universtiy headed by Prof. B. Siciliano, we are
designing and testing incrementally an architecture for robotic control based on active/expected
perception and steered by a planning module realized through our neural deductive system NSP
(Neuro Symbolic Processor) hardware implemented on FPGA processors.
In association with a local enterprise Neatec, we developed an intelligent system for the
detection of danger situations and for safety surveillance in railroad environments. This implies
detection and filtering of objects/people in images of the protected scene and requires both a
high degree of accuracy and dependability as well as very strict real time requirements. The
system has been developed in a WISARD neural architecture.
In the field of the Temporal Reasoning applied on QA we have implemented, with the ITC-
irst/Federico II collaboration, a Temporal Reasoning system for temporal expressions and we
have participated in the international competition “Time Expression Recognition and
Normalization” in order to evaluate this system.
In the field of cybernetics and neurosciences studies we have conceived a biologically inspired
visuo-motor control model based on a new interpretation for the functional role of mirror
neurons discovered by Rizzolatti and colleagues into the macaque’s F5 motor area. Our
functional interpretation includes an anticipatory mechanism, enabling one to verify whether the
actual, suitably coded visual input, matches an expected visual input computed on the basis of a
motor command sequence. We believe that this mirror neuron model (Mirror Expected
Perception model, MEP) should fit the biological data better than other models. This model has
been formalized in algorithmic terms and actual implementation is in progress.
A continuing line of analytical reflection on consciousness results in the proposal of a minimum
requirement for an algorithmic system which might hypothetically sustain consciousness.
Program for 2005
Within “Centro di Competenza sui Beni Culturali e Ambientali” we will complete the Oracle
implemented system with an agents architecture. The data base will encompass the Herculanum
and the Fusaro Columbarium data and will be put online.
The cooperation with the Robotic Group of the Faculty of Engineering will address the problem
of controlling the interaction of an industrial robotic arm both with a human operator and
another robotic arm in cooperative actions, making use of active/expected perception
techniques.
On the autonomous robotic line, the control based on active/expected perception and steered by
a planning module will be complemented by a network of fibred neurons designed to handle
analog signals from the sensors.
In the field of cybernetics and neurosciences studies we plan to implement the MEP model in a
simulated environment. Subsequently, we intend to probe MEP by embodying the proposed
visuo-motor control model into a real robotic platform interacting with another robot in a real
environment.