Transfer learning for deep object detectors in remotes sensing imaging datasets

Καλφαντής, Παύλος; Kalfantis, Pavlos

dc.contributor.author	Καλφαντής, Παύλος	el
dc.contributor.author	Kalfantis, Pavlos	en
dc.date.accessioned	2024-01-22T13:16:03Z
dc.date.available	2024-01-22T13:16:03Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/58628
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.26324
dc.description	Εθνικό Μετσόβιο Πολυτεχνείο--Μεταπτυχιακή Εργασία. Διεπιστημονικό-Διατμηματικό Πρόγραμμα Μεταπτυχιακών Σπουδών (Δ.Π.Μ.Σ.) "Επιστήμη Δεδομένων και Μηχανική Μάθηση"	el
dc.rights	Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα	*
dc.rights.uri	http://creativecommons.org/licenses/by-nc-nd/3.0/gr/	*
dc.subject	Remote Sensing	en
dc.subject	Μεταφορά Μάθησης	el
dc.subject	Ανίχνευση Αντικειμένων	el
dc.subject	Βαθιά Μάθηση	el
dc.subject	Μηχανική Μάθηση	el
dc.subject	Τηλεπισκόπηση	el
dc.subject	Object Detection	en
dc.subject	Transfer Learning	en
dc.subject	Deep Learning	en
dc.subject	Machine Learning	en
dc.title	Transfer learning for deep object detectors in remotes sensing imaging datasets	en
heal.type	masterThesis
heal.classification	Machine Learning	en
heal.language	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2023-07-12
heal.abstract	Object detection is a computer technology related to computer vision and image processing that deals with detecting instances of semantic objects of a certain class (such as humans, buildings, or cars) in digital images and videos. Object detection specifically in aerial and satellite images presents unique challenges compared to object detection in natural images, as aerial and satellite images often suffer from object size, scale and resolution variations, complex backgrounds and imbalanced datasets. Additionally, the large size of these images poses computational challenges for efficient and accurate object detection algorithms. Finally, the publicly available remote sensing imaging datasets are limited and creating and annotating new ones is a time and resource consuming endeavor. Transfer learning is a technique in machine learning in which knowledge learned from a task is re-used in order to boost performance on a related task. Reusing/transferring information from previously learned tasks to new tasks has the potential to significantly improve learning efficiency, as it allows the models to converge faster and potentially achieve better performance, even with a limited amount of data for the new task. In this thesis, the technique of transfer learning of pre-trained object detectors on a large dataset to new datasets with or without further training is investigated. The rationale behind this work concerns the study of the performance of the trained object detectors when evaluated on similar categories of new datasets, in order to identify the challenges and strengths of this approach as a solution to the challenges of object detection in satellite and aerial images. The current thesis is divided into three main parts. The first part of the thesis formulates the problem and challenges of object detection in remote sensing imagery, as well as it explains how transfer learning can be used to tackle these challenges. In addition, it investigates the current state-of-the-art object detection algorithms, where a variety of models are presented and five models are selected to be used in experiments throughout this thesis. In the next part of the thesis, the method that is followed during this project is formulated. The transfer learning approach that is followed is analyzed and the datasets that are being used in the experiments are presented. In addition, the metrics that are used as a basis for the evaluation of our experiments are explained. In the final part of the thesis, the results of the experiments are presented and analyzed. The results of training the object detectors in the baseline dataset, as well as the results of transferring the trained object detectors to two new datasets with similar objects are presented both quantitatively and qualitatively. Finally, conclusions regarding the method that was used and the results that were obtained are drawn and summarized.	en
heal.advisorName	Καράντζαλος, Κωνσταντίνος	el
heal.committeeMemberName	Στάμου, Γεώργιος	el
heal.committeeMemberName	Βουλόδημος, Αθανάσιος	el
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών	el
heal.academicPublisherID	ntua
heal.numberOfPages	73 σ.	el
heal.fullTextAvailability	false