TU Darmstadt / ULB / TUbiblio

Edge-related deep learning for semantic segmentation

Goebel, Mona ; Budde, Lina E. ; Iwaszczuk, Dorota
Hrsg.: Kersten, Thomas P. ; Tilly, Nora (2022)
Edge-related deep learning for semantic segmentation.
Dreiländertagung der DGPF, OVG und SGPF. Dresden, Germany (5.10.2022 - 6.10.2022)
doi: 10.24407/KXP:1796047864
Konferenzveröffentlichung, Bibliographie

Kurzbeschreibung (Abstract)

The identification of Land Use and Land Cover is an important task for creating maps or monitoring surface changes. The results in this work were obtained using a deep neural network called VGG19-Unet. Influenced by the increasing success of separately implementing edge-extracted inputs in a model, this work used the Canny method to create such edge extracted images. In the final model, one band was discarded from each of the original RGB and NIRRG images and replaced with an edge-extracted band, generating a closer connection between colour and edge bands. This model simultaneously learned eight classes. Segmentation results were found to be predominantly sharper when edge-extracted bands were included, and the model was more confident in its choice of classes. Lastly, improvements were achieved for objects with clear edges, such as buildings, as well as for objects with unclear edges, such as vegetation.

Typ des Eintrags: Konferenzveröffentlichung
Erschienen: 2022
Herausgeber: Kersten, Thomas P. ; Tilly, Nora
Autor(en): Goebel, Mona ; Budde, Lina E. ; Iwaszczuk, Dorota
Art des Eintrags: Bibliographie
Titel: Edge-related deep learning for semantic segmentation
Sprache: Englisch
Publikationsjahr: März 2022
Ort: Stuttgart
Verlag: Geschäftsstelle der DGPF
Reihe: Wissenschaftlich-Technische Jahrestagung der DGPF
Band einer Reihe: 42
Veranstaltungstitel: Dreiländertagung der DGPF, OVG und SGPF
Veranstaltungsort: Dresden, Germany
Veranstaltungsdatum: 5.10.2022 - 6.10.2022
DOI: 10.24407/KXP:1796047864
URL / URN: https://www.dgpf.de/src/tagung/jt2022/proceedings/start.html
Kurzbeschreibung (Abstract):

The identification of Land Use and Land Cover is an important task for creating maps or monitoring surface changes. The results in this work were obtained using a deep neural network called VGG19-Unet. Influenced by the increasing success of separately implementing edge-extracted inputs in a model, this work used the Canny method to create such edge extracted images. In the final model, one band was discarded from each of the original RGB and NIRRG images and replaced with an edge-extracted band, generating a closer connection between colour and edge bands. This model simultaneously learned eight classes. Segmentation results were found to be predominantly sharper when edge-extracted bands were included, and the model was more confident in its choice of classes. Lastly, improvements were achieved for objects with clear edges, such as buildings, as well as for objects with unclear edges, such as vegetation.

Freie Schlagworte: Remote Sensing, Photogrammetry, Deep Learning, Fernerkundung, Photogrammetrie, Geoinformationssystem
Fachbereich(e)/-gebiet(e): 13 Fachbereich Bau- und Umweltingenieurwissenschaften
13 Fachbereich Bau- und Umweltingenieurwissenschaften > Institut für Geodäsie
13 Fachbereich Bau- und Umweltingenieurwissenschaften > Institut für Geodäsie > Fernerkundung und Bildanalyse
Hinterlegungsdatum: 05 Aug 2022 06:49
Letzte Änderung: 06 Okt 2022 08:29
PPN: 497930838
Export:
Suche nach Titel in: TUfind oder in Google
Frage zum Eintrag Frage zum Eintrag

Optionen (nur für Redakteure)
Redaktionelle Details anzeigen Redaktionelle Details anzeigen