This thesis, developed in collaboration with Salvagnini Italia S.p.A., presents a deep learning-based approach for edge extraction from images depicting metal sheets placed within a laser cutting system. The goal is to estimate the position and orientation of each sheet by identifying its contours, allowing precise alignment of the laser head before the cutting process. The main problem that we address is achieving precise and robust edge detection under challenging lighting conditions and using limited computational resources. Several approaches presented in the literature were explored and compared, alongside preprocessing and postprocessing techniques, to identify the most suitable solution for the given constraints. Since inference time remained relatively high when processing high-resolution images, a knowledge distillation technique was used to transfer information from the main models to lighter and faster versions. Overall, this work highlights the trade-off between computational efficiency and precision in deep learning-based edge detection. Although traditional methods are faster, the proposed approach improves edge quality and robustness in visually challenging scenarios.
Questa tesi, sviluppata in collaborazione con Salvagnini Italia S.p.A., presenta un approccio basato sul deep learning per l’estrazione dei bordi da immagini rappresentanti fogli di lamiera posizionati all’interno di un sistema di taglio laser. L’obiettivo è stimare la posizione e l’orientamento di ciascun foglio identificandone i contorni, permettendo un allineamento preciso del puntatore laser prima del processo di taglio. Il problema principale affrontato consiste nel realizzare un’estrazione dei bordi precisa e robusta in condizioni di illuminazione complesse e utilizzando esclusivamente risorse computazionali limitate. Moltelici approcci presenti in letteratura sono stati analizzati e confrontati, insieme a tecniche di pre- e post-processing, al fine di individuare la soluzione più adatta ai vincoli imposti. Poiché il tempo di inferenza risultava relativamente elevato nell’elaborazione di immagini ad alta risoluzione, è stata applicata una tecnica di knowledge distillation per trasferire le informazioni dai modelli principali a delle versioni più leggere e veloci. In generale, questo lavoro evidenzia il compromesso tra efficienza computazionale e precisione nell’estrazione dei bordi basata sul deep learning. Sebbene i metodi tradizionali siano più veloci, l’approccio proposto migliora la qualità dei bordi e la robustezza in scenari visivamente sfidanti.
Deep learning edge extraction for laser cutting machine alignment
TOGNETTO, NICOLA
2024/2025
Abstract
This thesis, developed in collaboration with Salvagnini Italia S.p.A., presents a deep learning-based approach for edge extraction from images depicting metal sheets placed within a laser cutting system. The goal is to estimate the position and orientation of each sheet by identifying its contours, allowing precise alignment of the laser head before the cutting process. The main problem that we address is achieving precise and robust edge detection under challenging lighting conditions and using limited computational resources. Several approaches presented in the literature were explored and compared, alongside preprocessing and postprocessing techniques, to identify the most suitable solution for the given constraints. Since inference time remained relatively high when processing high-resolution images, a knowledge distillation technique was used to transfer information from the main models to lighter and faster versions. Overall, this work highlights the trade-off between computational efficiency and precision in deep learning-based edge detection. Although traditional methods are faster, the proposed approach improves edge quality and robustness in visually challenging scenarios.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tognetto_Nicola.pdf
Accesso riservato
Dimensione
34.33 MB
Formato
Adobe PDF
|
34.33 MB | Adobe PDF |
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/20.500.12608/98774