Sustav za snimanje radarom sa sintetičkom aperturom i integriranom klasifikacijom objekata

doctoral thesis

• Title (english): System for synthetic aperture radar imaging with integrated object classification
• Author: Filip Turčinović
• Mentor: Marko Bosiljevac (mentor)
• Committee member: Sonja Grgić (predsjednik povjerenstva)
• Committee member: Irena Galić (član povjerenstva)
• Committee member: Gordan Šišul (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2024-06-10, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing
• Universal decimal classification (UDC): 004 - Computer science and technology. Computing. Data processing

Abstract

Disertacija prati razvoj sustava za snimanje radarom sa sintetičkom aperturom i klasifikacijom objekata. Mikroračunalo oko kojeg je sustav implementiran provodi funkcije radarskog snimanja, rekonstrukcije radarske slike i modela dubokog učenja. Radarsko snimanje radi se po principu zemaljskog radara sa sintetičkom aperturom (GBSAR), a rekonstrukcija slike algoritmom Omega-K. Za klasifikaciju objekata uspoređene su različite arhitekture modela, od kojih se istaknula arhitektura rezidualne mreže (ResNet). Analizirani su rezultati modela treniranih sirovim radarskim podacima s rezultatima modela treniranih rekonstruiranim slikama. Pokazano je da, iako je lakše interpretirati rekonstruirane slike, modeli temeljeni na neobrađenim (sirovim) podacima postižu bolje rezultate. Sirovim podacima je prilagođena arhitektura rezidualne mreže s 18 slojeva (ResNet-18). Sustav povećava energetsku učinkovitost komunikacijom između klasifikacijskog modela i radara. Pokazano je da dinamičkim postavljanjem radarskih parametara 73% detaljnijih mjerenja nije potrebno provoditi za postizanje visoke razine točnosti klasifikacije. Istražena je mogućnost dodavanja novog izvora informacije modelu kroz snimanje dvjema linearnim polarizacijama. Opisani su pristupi povezivanja matrica dobivenih horizontalnom i vertikalnom polarizacijom te arhitektura modela koji te matrice prima odvojeno i spaja vektore značajki prije potpuno povezanog sloja. Kroz iteracije je pokazano da dodatna informacija iz matrice snimljene drugom polarizacijom povećava točnost modela.

Abstract (english)

This dissertation explores the development of a synthetic aperture radar (SAR) imaging and object classification system. Built around a microcomputer, the system integrates radar imaging, image reconstruction, and deep learning model. The radar operates on the Ground-Based Synthetic Aperture Radar (GBSAR) principle, with image reconstruction using the Omega-K algorithm. Various model architectures were compared for object classification, with the Residual Network (ResNet) architecture standing out. The dissertation analyzes models trained on raw radar data versus those trained on reconstructed images. Although reconstructed images are more intuitive, models based on raw data achieve superior results. The ResNet-18 architecture was adapted for raw data, maintaining the horizontal dimension of the data matrix during training. The system enhances energy efficiency using communication between the classification model and the radar. Dynamic adjustment of radar parameters reduces the need for detailed measurements by 73% while maintaining high classification accuracy. The potential of incorporating additional information to the model based on dual linear polarization imaging was also explored. Approaches for combining matrices recorded using horizontal and vertical polarizations were described, as well as a model architecture that processes these matrices separately before merging feature vectors before fully connected layer. Iterative testing demonstrated that additional information obtained using matrices recorded with other polarization improves model accuracy.

• Keywords: radar sa sintetickom aperturom; klasifikacija objekata; ResNet-18; rekonstrukcija radarske slike; algoritam Omega-K; polarizacija elektromagnetskog vala
• Keywords (english): Ground Based Synthetic Aperture Radar; Object Classification; ResNet18; Radar Image Reconstruction; Omega-K Algorithm; EM Wave Polarization
• Language: croatian
• URN:NBN: urn:nbn:hr:168:296160
• Promotion: 2024
• Study programme: Title: Computer Science; Study programme type: university; Study level: postgraduate; Academic / professional title: Doktor znanosti (Doktor znanosti)
• Type of resource: Text
• Extent: 89 str.
• File origin: Born digital
• Access conditions: Open access
• Created on: 2024-08-29 07:59:14