Corrosion detection with computer vision and deep learning

Ματθαίου, Αχιλλέας; Matthaiou, Achilleas

dc.contributor.author	Ματθαίου, Αχιλλέας	el
dc.contributor.author	Matthaiou, Achilleas	en
dc.date.accessioned	2021-03-30T09:41:21Z
dc.date.available	2021-03-30T09:41:21Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/53194
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.20892
dc.rights	Default License
dc.subject	Corrosion Detection	en
dc.subject	Deep learning	en
dc.subject	Computer vision	en
dc.subject	Object detection	en
dc.subject	Convolutionalnneural network	en
dc.subject	Βαθιά μάθηση	el
dc.subject	Εντοπισμός διάβρωσης	el
dc.subject	Όραση υπολογιστών	el
dc.subject	Νευρωνικά δίκτυα	el
dc.subject	Εντοπισμός αντικειμένου	el
dc.title	Corrosion detection with computer vision and deep learning	en
heal.type	bachelorThesis
heal.classification	Artificial intelligence	en
heal.classification	Marine structures monitoring	en
heal.classification	Deep learning	en
heal.classification	Computer vision	en
heal.language	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2021-03-08
heal.abstract	This study investigates and test solutions for automated corrosion detection processes that focus on the visual attributes of corrosion. Artificial intelligence methods are employed to extract information from images. Corroded surfaces have two visually identified attributes color and texture. To detect corrosion based on color, a color tracking algorithm is created and tested using images from different compartments of vessels. To detect corrosion based on texture, deep learning algorithms are used, and two approaches are tested. The first approach is a binary classification model trained using a Convolutional Neural Network (CNN) architecture employing transfer learning. The model is also used by a sliding algorithm to allow detection and localization in large corroded plates. The second approach treats corrosion detection as an object detection problem. A Single Shot Detector (SSD) is trained using transfer learning to detect corrosion on real world images. To support training and testing of all models two datasets are created. The first dataset consists of images of metals corroded in a laboratory environment, while the second dataset from real world images of corroded compartments from bulk carriers’ inspections. The study finds all three methods capable to perform corrosion detection with the deep learning approaches yielding better results. Comparing the two deep learning approaches, object detection is found to be more suitable for real world examples.	en
heal.advisorName	Παπαλάμπρου, Γεώργιος	el
heal.committeeMemberName	Ζαραφωνίτης, Γεώργιος	el
heal.committeeMemberName	Σαμουηλίδης, Μανόλης	el
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Ναυπηγών Μηχανολόγων Μηχανικών. Τομέας Ναυτικής Μηχανολογίας. Εργαστήριο Ναυτικής Μηχανολογίας	el
heal.academicPublisherID	ntua
heal.numberOfPages	97 σ.	el
heal.fullTextAvailability	false