In this thesis, we tackle the problem of the localization of highly reflective metallic objects inside a laser cutting machine. We propose a fully autonomous end-to-end computer vision application that achieves the goal and that improves the performance of current solutions. Thanks to recent advancements in deep learning and computer vision, repetitive and tedious tasks that require human visual perception supervision can potentially be delegated to Artificial Intelligence (AI) algorithms. But these techniques often tend to remain relegated to academic research and are not considered reliable enough for industrial applications: instead, our algorithm was deployed as part of an artificial vision system on laser cutting machines. Our solution is based on the DeepLab v3 CNN model with a custom loss function that minimizes the error near the edges of objects and satisfies the requirements for fast and precise localization of highly reflective metallic objects. We compare our results with other state of art models and analyze the trade-off. Moreover, the proposed system is compared against other localization strategies that are currently installed on the machine.
In questa tesi affrontiamo il problema della localizzazione di oggetti metallici altamente riflettenti all'interno di una macchina da taglio laser. Proponiamo un'applicazione di visione artificiale end-to-end completamente autonoma che raggiunge l'obiettivo e che migliora le prestazioni delle soluzioni attuali. Grazie ai recenti progressi nell' deep learning e nella visione artificiale, compiti ripetitivi e noiosi che richiedono la supervisione della percezione visiva umana possono essere potenzialmente delegati ad algoritmi di Intelligenza Artificiale (AI). Ma queste tecniche spesso tendono a rimanere relegate alla ricerca accademica e non sono considerate sufficientemente affidabili per applicazioni industriali: il nostro algoritmo è stato invece implementato come parte di un sistema di visione artificiale su macchine da taglio laser. La nostra soluzione si basa sul modello CNN DeepLab v3 con una loss function personalizzata che riduce al minimo l'errore vicino ai bordi degli oggetti e soddisfa i requisiti per una localizzazione rapida e precisa di oggetti metallici altamente riflettenti. Confrontiamo i nostri risultati con altri modelli all'avanguardia e analizziamo i compromessi. Inoltre, il sistema proposto viene confrontato con altre strategie di localizzazione attualmente installate sulla macchina.
A deep learning-based approach for detection and segmentation of reflective materials
COLPO, CRISTIANO
2021/2022
Abstract
In this thesis, we tackle the problem of the localization of highly reflective metallic objects inside a laser cutting machine. We propose a fully autonomous end-to-end computer vision application that achieves the goal and that improves the performance of current solutions. Thanks to recent advancements in deep learning and computer vision, repetitive and tedious tasks that require human visual perception supervision can potentially be delegated to Artificial Intelligence (AI) algorithms. But these techniques often tend to remain relegated to academic research and are not considered reliable enough for industrial applications: instead, our algorithm was deployed as part of an artificial vision system on laser cutting machines. Our solution is based on the DeepLab v3 CNN model with a custom loss function that minimizes the error near the edges of objects and satisfies the requirements for fast and precise localization of highly reflective metallic objects. We compare our results with other state of art models and analyze the trade-off. Moreover, the proposed system is compared against other localization strategies that are currently installed on the machine.File | Dimensione | Formato | |
---|---|---|---|
Colpo_Cristiano.pdf
accesso riservato
Dimensione
4.92 MB
Formato
Adobe PDF
|
4.92 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/20.500.12608/35223