The present PhD thesis covers a research about real-time spectral classification exploiting machine learning tools and in particular deep learning with neural networks. Spectroscopy is the study of the interaction between matter and electromagnetic radiation. In particular, we focused on Raman spectroscopy because it allows a chemical compound identification that has many fields of application. Our approach to solve spectral classification is a combination of signal processing methods and machine learning tools. First, it has been analysed the most important spectrum feature that could affect classification accuracy: signal-to-noise ratio. Since the level of noise must be reduced both at acquisition time and at processing time, an investigation of the noise sources in a CCD device has been carried on in order to identify the main parameters that concur to generate noise. Once that the best acquisition strategy has been demonstrated, we focused on a Cultural Heritage case of study about pigment spectra classification, exploiting machine learning tools to overcome the problem of a limited reference database. A hierarchical pipeline of signal processing methods has been designed to implement the necessary Data Augmentation. In addition, these methods have been enhanced with the use of Generative Adversarial Networks (GANs). Once that an augmented dataset has been generated, we tested it on 3 different neural network architectures, finding that the Convolutional Neural Networks (CNN) with GAN enhancement is the best approach for classification. The implemented technique of the present thesis is potentially applicable in every spectroscopic analysis that lacks a sufficient number of training reference examples.

Real-Time Algorithms for Spectral Classification Tasks / DI FRISCHIA, Stefano. - (2021 Jun 14).

Real-Time Algorithms for Spectral Classification Tasks

DI FRISCHIA, STEFANO
2021-06-14T00:00:00+02:00

Abstract

The present PhD thesis covers a research about real-time spectral classification exploiting machine learning tools and in particular deep learning with neural networks. Spectroscopy is the study of the interaction between matter and electromagnetic radiation. In particular, we focused on Raman spectroscopy because it allows a chemical compound identification that has many fields of application. Our approach to solve spectral classification is a combination of signal processing methods and machine learning tools. First, it has been analysed the most important spectrum feature that could affect classification accuracy: signal-to-noise ratio. Since the level of noise must be reduced both at acquisition time and at processing time, an investigation of the noise sources in a CCD device has been carried on in order to identify the main parameters that concur to generate noise. Once that the best acquisition strategy has been demonstrated, we focused on a Cultural Heritage case of study about pigment spectra classification, exploiting machine learning tools to overcome the problem of a limited reference database. A hierarchical pipeline of signal processing methods has been designed to implement the necessary Data Augmentation. In addition, these methods have been enhanced with the use of Generative Adversarial Networks (GANs). Once that an augmented dataset has been generated, we tested it on 3 different neural network architectures, finding that the Convolutional Neural Networks (CNN) with GAN enhancement is the best approach for classification. The implemented technique of the present thesis is potentially applicable in every spectroscopic analysis that lacks a sufficient number of training reference examples.
Real-Time Algorithms for Spectral Classification Tasks / DI FRISCHIA, Stefano. - (2021 Jun 14).
File in questo prodotto:
File Dimensione Formato  
Tesi Dottorato di Ricerca_ICT_Di Frischia.pdf

accesso aperto

Descrizione: Tesi
Tipologia: Tesi di dottorato
Dimensione 12.35 MB
Formato Adobe PDF
12.35 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: http://hdl.handle.net/11697/177860
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact