Development of non-intrusive solutions for identification (THID) and authentication in the THz domain

  Defense of doctoral thesis by Dragos-Florin NASTASIU, for the  University  Grenoble Alpes, speciality " SIGNAL IMAGE PAROLE TELECOMS"

Keywords : 
THz Spectroscopy,THz Tag,Identification and Authentication Techniques,Spectral Analysis,Transient Signal Analysis,Machine Learning

Abstract :
THz imaging is an emerging field since the technological advances in terms of THz radiation emission and detection equipment. The main objective of the thesis is to contribute and to improve THz imaging systems, from image reconstruction and analysis to image classification tasks. In the first part of the thesis, we tackle the amplitude estimation challenge under ideal and multiplicative noise conditions. The multiplicative noise deforms the phase and introduces complex artefacts, such as contour information loss and contrast degradation, that cannot be eliminated using state-of-the-art image reconstruction techniques. In this regard, we introduce five novel reconstruction methods which exploit the phase diagram representation of signals. Two of the methods are based on phase-diagram match filtering to estimate the amplitude in both conditions. Another two methods use the concept of dynamic time warping (DTW) to increase the capability to model the multiplicative type of noise. Lastly, we exploit the dynamic of the phase trajectory described by the curvatures to reconstruct the image. From the large pool of methods, we evaluate throughout the thesis that the curvature-based method efficiently reconstructs the image in both ideal and noisy contexts. After an efficient image reconstruction, the second part of the thesis, we study image analysis and classification methods considering the instabilities of real-world imaging systems, such as translations and rotations. In this sense, we propose to use translation and rotation invariant wavelet packet decompositions, that provide a unique and optimal representation of an image, regardless if the image is translated or rotated. Based on the invariant image representations, novel feature extraction techniques are introduced such as vertical, horizontal, N-directional and N-zonal frameworks. Additionally, two feature structures are introduced and that consider the frequency partitioning of the wavelet decomposition and are adapted to work with Graph Neural Networks (GNNs) and classic ML classifiers such as k-nearest neighbors (k-NN), support vector machine (SVM), etc. Overall, our proposed approaches increase the accuracy of all classifiers.

Jury members :

• Frédéric GARET, Professor of universities - University of Savoie Mont-Blanc : Supervisor
• Andrei ANGHEL, National Professor - University of Science and Technology POLITEHNICA : Reviewer
• Emmanuel TROUVE, Professor of universities - LISTIC – Polytech Annecy-Chambéry : Examiner
• Cornel IOANA, Assistant professor -  University Grenoble Alpes : CoSupervisor
• Patrick MOUNAIX,  Research Director - CNRS : Reviewer
• Alexandru SERBANESCU,  Professor  Military Technical Academy Ferdinand I : CoSupervisor
• Sri  KRISHNAN,  Professor -Faculty of Engineering & Architectural Science (FEAS) : Examiner
• Alexandre LOCQUET, Research manager - CNRS : Examiner