Currently, the analysis of road continuity in satellite images is a complex challenge due to the difficulty in identifying the directional vector of road sections, especially when the satellite view of roads is obstructed by trees or other structures. Nowadays, most research focuses on optimizing the deep learning networks already available for this purpose, however, the segmentation accuracy of these is especially affected by the loss function used in testing. Currently, little research has been published on loss functions used in road segmentation. To solve this problem, the study focused specifically on a loss function called GapLoss, which can be combined with any proposed segmentation network. The procedure of this loss function is purely a topological pixel analysis and is based on five steps, at the end of which a higher weight is given for problematic pixels representing non-real road breaks. In this study, three semantic segmentation models will be examined to conduct comparisons of the results obtained using a dataset. The purpose is to compare the values obtained from this loss function and the alternative version, with those obtained from the Dice Loss based model, used as an objective. Also in the conclusions it should be checked whether in the predicted image, the road continuity is continuous enough, especially at intersections and parts where roads are hidden by obstacles of various types.
Attualmente l’analisi della continuità stradale nelle immagini satellitari è una sfida complessa dovuta alla difficoltà nell’individuare il vettore direzionale dei tratti stradali, soprattutto quando la vista satellitare delle strade è ostruita da alberi o altre strutture. Al giorno d’oggi, la maggioranza delle ricerche si sofferma sull'ottimizzazione delle reti deep learning già disponibili per questo scopo, tuttavia, l'accuratezza della segmentazione di queste è influenzata specialmente dalla loss function usata nei test.\\ Attualmente sono state pubblicate poche ricerche su loss function usate nell’ambito della segmentazione stradale. Per risolvere questo problema, lo studio si è concentrato in particolare su una loss function denominata GapLoss, combinabile con qualsiasi rete di segmentazione proposta. Il procedimento di questa loss function è prettamente un'analisi topologica dei pixel ed è basato su cinque fasi, alla fine delle quali si attribuisce un peso maggiore per i pixel problematici che rappresentano interruzioni stradali non reali. \\ In questo studio verranno esaminati tre modelli di segmentazione semantica per condurre confronti tra i risultati ottenuti utilizzando un dataset. Lo scopo è confrontare i valori ottenuti da questa loss function e dalla versione alternativa, con quelli ottenuti dal modello basato su Dice Loss, fissati come obbiettivo. Inoltre nelle conclusioni va controllato se nell'immagine predetta, la continuità stradale risulta abbastanza continua, soprattutto nelle intersezioni e nelle parti in cui le strade sono nascoste da ostacoli di vario tipo.
GapLoss: una loss function per segmentazione semantica di strade in immagini satellitari
MARCON, KRISTIAN
2021/2022
Abstract
Currently, the analysis of road continuity in satellite images is a complex challenge due to the difficulty in identifying the directional vector of road sections, especially when the satellite view of roads is obstructed by trees or other structures. Nowadays, most research focuses on optimizing the deep learning networks already available for this purpose, however, the segmentation accuracy of these is especially affected by the loss function used in testing. Currently, little research has been published on loss functions used in road segmentation. To solve this problem, the study focused specifically on a loss function called GapLoss, which can be combined with any proposed segmentation network. The procedure of this loss function is purely a topological pixel analysis and is based on five steps, at the end of which a higher weight is given for problematic pixels representing non-real road breaks. In this study, three semantic segmentation models will be examined to conduct comparisons of the results obtained using a dataset. The purpose is to compare the values obtained from this loss function and the alternative version, with those obtained from the Dice Loss based model, used as an objective. Also in the conclusions it should be checked whether in the predicted image, the road continuity is continuous enough, especially at intersections and parts where roads are hidden by obstacles of various types.File | Dimensione | Formato | |
---|---|---|---|
Marcon_Kristian.pdf
accesso aperto
Dimensione
3.29 MB
Formato
Adobe PDF
|
3.29 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/38007