The following thesis develops around a real problem of identification and classification of a superficial defect on a bar of stainless steel. The main objective is to evaluate the performance of a 2D vision system based on convolutional neural networks (Faster R-CNN architecture), to understand its real ability to replace or support the human operator in quality control. To achieve this goal, a statistical analysis was conducted to measure the concordance and repeatability of the system with respect to traditional inspection. The adopted method, the experimental setup, and the characteristics of the machine learning model used are described. In addition to the validation phase, some optimizations aimed at improving overall performance were introduced and tested. The results are presented and discussed critically, highlighting both the effects of the corrections introduced and the limits that are still present. The thesis concludes with a reflection on the future developments needed to overcome these limits and move closer to the full automation of the quality control process.
La seguente tesi si sviluppa attorno ad un problema reale di identificazione e classificazione di difetti superficiali su barre di acciaio inossidabile. L’obiettivo principale è valutare le prestazioni di un sistema di visione 2D basato su reti neurali convoluzionali (architettura Faster R-CNN), per capirne la reale capacità di sostituire o affiancare l’operatore umano nel controllo qualità. Per raggiungere questo scopo è stata condotta un’analisi statistica finalizzata a misurare la concordanza e la ripetibilità del sistema rispetto all’ispezione tradizionale. Vengono descritti il metodo adottato, il setup sperimentale e le caratteristiche del modello di apprendimento automatico utilizzato. Oltre alla fase di validazione, sono state introdotte e testate alcune ottimizzazioni mirate al miglioramento delle prestazioni complessive. I risultati ottenuti sono presentati e discussi in modo critico, mettendo in evidenza sia gli effetti delle correzioni introdotte sia i limiti tuttora presenti. La tesi si conclude con una riflessione sugli sviluppi futuri necessari per superare tali limiti e avvicinarsi alla piena automazione del processo di controllo qualità.
Implementazione di un sistema di rilevazione ottica di difetti su barre in acciaio inossidabile basato su algoritmi di intelligenza artificiale
DALLA BONA, ENRICO
2024/2025
Abstract
The following thesis develops around a real problem of identification and classification of a superficial defect on a bar of stainless steel. The main objective is to evaluate the performance of a 2D vision system based on convolutional neural networks (Faster R-CNN architecture), to understand its real ability to replace or support the human operator in quality control. To achieve this goal, a statistical analysis was conducted to measure the concordance and repeatability of the system with respect to traditional inspection. The adopted method, the experimental setup, and the characteristics of the machine learning model used are described. In addition to the validation phase, some optimizations aimed at improving overall performance were introduced and tested. The results are presented and discussed critically, highlighting both the effects of the corrections introduced and the limits that are still present. The thesis concludes with a reflection on the future developments needed to overcome these limits and move closer to the full automation of the quality control process.| File | Dimensione | Formato | |
|---|---|---|---|
|
DallaBona_Enrico.pdf
Accesso riservato
Dimensione
7.73 MB
Formato
Adobe PDF
|
7.73 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/20.500.12608/100033