Content replacement and inpainting in sections of remote sensing images using deep learning techniques.

Drimiskianakis, Zacharias; Δριμισκιανάκης, Ζαχαρίας

dc.contributor.author	Drimiskianakis, Zacharias	en
dc.contributor.author	Δριμισκιανάκης, Ζαχαρίας	el
dc.date.accessioned	2025-03-19T09:46:57Z
dc.date.available	2025-03-19T09:46:57Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/61359
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.29055
dc.rights	Αναφορά Δημιουργού - Μη Εμπορική Χρήση - Παρόμοια Διανομή 3.0 Ελλάδα	*
dc.rights.uri	http://creativecommons.org/licenses/by-nc-sa/3.0/gr/	*
dc.subject	Computer vision	en
dc.subject	Image inpainting	en
dc.subject	Remote sensing	en
dc.subject	Deep learning	en
dc.subject	Diffusion models	en
dc.subject	sdxl for inpainting	en
dc.subject	Textual inversion	en
dc.subject	lora	en
dc.subject	Dreambooth	en
dc.subject	Όραση υπολογιστών	el
dc.subject	Ενδοσυμπλήρωση εικόνας	el
dc.subject	Τηλεπισκοπικές εικόνες	el
dc.subject	Βαθιά μάθηση	el
dc.subject	Μοντέλα διάχυσης	el
dc.subject	Κειμενική αντιστροφή	el
dc.title	Content replacement and inpainting in sections of remote sensing images using deep learning techniques.	en
dc.title	Αντικατάσταση περιεχομένου και ενδοσυμπλήρωση σε τμήματα τηλεπισκοπικών εικόνων με τεχνικές βαθιάς μάθησης.	el
heal.type	bachelorThesis
heal.generalDescription	Η παρούσα εργασία εστιάζει στην ενδοσυμπλήρωση τηλεπισκοπικών εικόνων υψηλής ανάλυσης χρησιμοποιώντας προηγμένες μεθόδους βαθιάς μάθησης, με ιδιαίτερη έμφαση στα μοντέλα διάχυσης.	el
heal.classification	Remote sensing	en
heal.classification	Τηλεπισκόπηση	el
heal.language	el
heal.language	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2024-10-16
heal.abstract	The present bachelor thesis focuses on the inpainting of high-resolution remote sensing images using advanced deep learning methods, with particular emphasis on diffusion models. Initially, a theoretical background on inpainting methods is provided, which is explored through classical computer vision techniques and modern deep learning approaches. Special attention is given to diffusion models, explaining their structure, training methods, sampling procedures, and advanced variations. Next, an initial methodology and corresponding execution examples are presented, implemented to examine the capability of the chosen diffusion model to handle various scenarios of remote sensing image inpainting. Based on the conclusions and limitations of the aforementioned methodology, a new methodology is introduced, along with execution examples, focusing on localized content replacement. Moreover, tests for inpainting were conducted on the NIR channel, along with attempts for guided image generation using a ControlNet model and fine-tuning tests of the pre-trained model on specific remote sensing entities using methods such as Textual Inversion and Dreambooth. In the final conclusions and recommendations for improvement, the strengths and limitations of the approach are highlighted, with suggestions for enhancing the accuracy and efficiency of inpainting methods on high-resolution remote sensing images.	en
heal.sponsor	Hellenic Military Geographical Service	en
heal.advisorName	Karantzalos, Konstantinos	en
heal.committeeMemberName	Karathanasi, Vasilia	en
heal.committeeMemberName	Papoutsis, Ioannis	en
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Αγρονόμων και Τοπογράφων Μηχανικών. Τομέας Τοπογραφίας. Εργαστήριο Τηλεπισκόπησης	el
heal.academicPublisherID	ntua
heal.numberOfPages	139 σ.	el
heal.fullTextAvailability	false
heal.fullTextAvailability	false