1 A neural approach to the analysis of CHIMERA experimental data CHIMERA Collaboration S.Aiello 1, M. Alderighi 2,3, A.Anzalone 4, M.Bartolucci 5, G.Cardella 1, S.Cavallaro 4,7, M. DAgostino 6,E.DeFilippo 1, E.Geraci 4, M.Geraci 1, F.Giustolisi 4,7, P.Guazzoni 3,5, M.Iacono Manno 4, G.Lanzalone 1,7, G.Lanzanò 1, S.LoNigro 1,7, G.Manfredi 5, A.Pagano 1, M.Papa 1, S.Pirrone 1, G.Politi 1,7, F.Porto 4,7, S.Russo 5, S.Sambataro 1,7, G.Sechi 2,3, L.Sperduto 4,7, C.Sutera 1, L.Zetta 3,5 1 Istituto Nazionale di Fisica Nucleare, sez di Catania, Catania, Italy 2 Istituto di Fisica Cosmica, CNR, Milano, Italy 3 Istituto di Fisica Nucleare, sez. di Milano, Milano, Italy 4 Istituto di Fisica Nucleare, Laboratorio Nazionale del Sud, Catania, Italy 5 Dipartimento di Fisica dellUniversita, Milano, Italy 6 Dipartimento di Fisica dell'Universita degli Studi and Istituto di Fisica Nucleare, sez. di Bologna, Bologna,Italy 7 Dipartimento di Fisica dell'Universita, Catania, Italy
2 Outline Detector characteristics Automatic data analysis Proposed approaches Our neural approach System overview Results
3 CHIMERA (Charged Heavy Ion Mass and Energy Resolving Array) 1192 Si-CsI(TI) detection cells 9 wheels
4 Preamplifier Photodiode Silicon detector CsI(TI) detector TOF E FastSlow Fast Slow Fast E - Si Detection cell
5 58 Ni + 27 Al E inc = 30 AMev Scatter plot from CHIMERA sparse data low S/N density variation –high frequency: noise –characteristic frequency: ridges/valleys –low frequency: background
6 E-Si Fast-CsI(TI) banana extraction ? E-Si Counts
7 Fast-CsI(TI) E-Si Fast-CsI(TI) E-Si Counts 1-D frequency distribution Z-lines
8 Proposed approach FFT not satisfactory results filtering edge detection = ill-posed problem contextual image segmentation [Benkirane et al. 95]: Canny filtering + a priori information not easily applicable interactive technique unpractical for a lot of spectra yet, density modulation can be easily perceived by sight
9 Our solution Using emergent perception mechanisms of biological visual systems Grossbergs neural networks mathematically defined extract information from the global structure of data (rather local relationship) no training successfully applied to SAR and satellite images (noisy and incomplete)
10 Implementation 2 levels of neural networks for cluster determination Procedural algorithms for frequency distribution construction Matlab (PC Pentium II, 400MHz) pixel processing windows
11 Neural system Window ADD net BF net Level 2: oriented completion (Bipole Filter) Level 1: Adaptive Density Discrimination Input
12 Level 1: ADD net on-center off-surround shunting network density information processing –comparison between on-center and off-surround areas –low-pass filtering of the spatial frequencies in the input window sensitivity to ridge-valley modulation clusters as incomplete and irregular strips Input CENTER SURROUND
13 ADD net input window on-center convolution off-surround convolution inhibitory input excitatory input + - i i i j j j x ij
14 Level 2: BF nets additive networks long-term cooperation along selected directions –bipole filters –different filtering masks according to hyperbolic trends of data clusters as complete strips 105° 135° Ex.
15 Example 1 valley clusters
16 ridge clusters Example 2
17 Conclusions Grossbergs approach is good for automatic determination of bananas Density processing is –dependent on the image structure only –independent from the underlying physics Intensive computation ( neurons) Processing whole matrices and improving algorithm efficiency as future works