Cosmic ray tomography based object reconstruction and recognition
Date
2023-10-11
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Tehnoloogia kiire areng ja valdkonnas tehtud edukad teadustööd on toonud endaga kaasa tehisintellektil põhinevate meetodite kasutamise laialdase leviku madalamatasemeliste ülesannete, näiteks objektituvastuse, lahendamisel. Objektituvastus proovib leida huvipakkuvaid objekte pildilt või pildijadadelt. Tuvastuse käigus lokaliseeritakse ja klassifitseeritakse pildil olevad objektid.
Objektituvastus leiab laialdast kasutust mitmete reaalmaailma probleemide lahendamisel ja on seetõttu üks peamisi ülesandeid, mida arvutinägemises lahendatakse. Seda rakendatakse näiteks tomograafiasüsteemide väljundite uurimises, inimkäitumise analüüsis, meditsiinis ja spordi valdkonnas.
Vaatamata objektituvastuse pikaajalisele käibel olemisele, on selles valdkonnas jätkuvalt palju väljakutsuvaid teemasid, mida uurida. Üheks selliseks näiteks on süvaõppe kasutamine müüontomograafias. Müüonid on elektroni sarnased elementaarosakesed, mis erinevaid materjale läbides neelduvad ja hajuvad erineva nurga all. Need hajumisnurgad peidavad endas informatsioon läbitud objektide kohta. Maal tekivad müüonid enamasti kosmilise kiirguse sisenemisel Maa atmosfääri. Müüontomograafias mõõdetakse osakeste sisenemis- ja väljumisnurkasid ümber huvipakkuva objekti. Nende nurkade põhjal rekonstrueeritakse objektist tomograafilised pildid. Saadud pilte kasutatakse objektituvastuses. Objektituvastuse tulemus on otseselt mõjutatud rekonstruktsioonist ja sellest kui täpselt suudetakse lokaliseerida müüonite tabamuspunkte detektori plaatidelt.
Sügavate närvivõrkude kasutamine ülesande lahendamisel parandab märkimisväärselt tulemuste täpsust võrreldes traditsioonilise raskuskeskme meetodiga. Käesolevas töös näidatakse sügavate konvolutsiooniliste närvivõrkude edukat utiliseerimist objektituvastuse ülesannete lahendamisel keerukate ja mitmekesiste andmekogumite peal.
Improvement of the technology and successful research in the artificial intelligence (AI) field have resulted in the surge of using AI-based methods to handle low-level tasks such as object recognition. Object recognition aims to find the objects within an image or sequence of images. The goal is to localize the existing objects within the image and identify the classes of them. Due to its wide application in significant real-world problems, object recognition is one of the main tasks in computer vision. Some examples of these applications include tomography systems, human-behaviour analysis, medical imaging, and sports. Although object recognition has been a hot topic for quite some time, there are still many challenges remaining due to the wide range of application areas. For example, such emerging technologies as cosmic ray tomography have a big research gap for utilizing deep learning techniques to improve the system’s object recognition performance. In order to improve the reconstruction and recognition performance in a cosmic ray tomography system, a neural network based approach is presented in this thesis. A cosmic ray tomography system uses the hit positions of muons on detector plates to reconstruct the images of volume of interest (VOI). Afterwards, these images are used to recognize the objects. Therefore, the recognition performance is directly impacted by the reconstruction performance, hence also by the muon hit position estimation. A significant improvement over the conventional Center of Gravity (CoG) method is obtained by the utilisation of deep neural networks (DNN) for this task. Moreover, in this thesis, object recognition methods based on deep convolutional neural networks (DCNN) are presented. Their capabilities to recognize the objects from a single image and a time-series data are presented. Successful performances are demonstrated through the experimentation with challenging tasks on the diverse datasets.
Improvement of the technology and successful research in the artificial intelligence (AI) field have resulted in the surge of using AI-based methods to handle low-level tasks such as object recognition. Object recognition aims to find the objects within an image or sequence of images. The goal is to localize the existing objects within the image and identify the classes of them. Due to its wide application in significant real-world problems, object recognition is one of the main tasks in computer vision. Some examples of these applications include tomography systems, human-behaviour analysis, medical imaging, and sports. Although object recognition has been a hot topic for quite some time, there are still many challenges remaining due to the wide range of application areas. For example, such emerging technologies as cosmic ray tomography have a big research gap for utilizing deep learning techniques to improve the system’s object recognition performance. In order to improve the reconstruction and recognition performance in a cosmic ray tomography system, a neural network based approach is presented in this thesis. A cosmic ray tomography system uses the hit positions of muons on detector plates to reconstruct the images of volume of interest (VOI). Afterwards, these images are used to recognize the objects. Therefore, the recognition performance is directly impacted by the reconstruction performance, hence also by the muon hit position estimation. A significant improvement over the conventional Center of Gravity (CoG) method is obtained by the utilisation of deep neural networks (DNN) for this task. Moreover, in this thesis, object recognition methods based on deep convolutional neural networks (DCNN) are presented. Their capabilities to recognize the objects from a single image and a time-series data are presented. Successful performances are demonstrated through the experimentation with challenging tasks on the diverse datasets.
Description
Väitekirja elektrooniline versioon ei sisalda publikatsioone
Keywords
cosmic rays, pattern recognition, muon tomography, artificial neural networks