Template Matching is a fundamental technology in industrial automation, essential for tasks like quality control and robotic navigation. However, most algorithms struggle to detect common industrial symbols. In fact, these symbols are intentionally designed with simple shapes and plain colors for human readability, which provides little information for conventional Computer Vision methods. While Deep Learning offers a valid solution, its dependence on large datasets, long training, and expensive hardware makes it impractical for many dynamic industrial environments. This thesis presents a novel, training-free method that turns this challenge into an opportunity. Our expert-driven system exploits the unique characteristics of the templates by combining their known color and geometric properties. The algorithm first uses color information to efficiently process the image for potential candidates and then analyzes their shape to achieve precise and robust detection. Experimental results demonstrate that our approach achieves state-of-the-art accuracy, outperforming established methods, including a Deep Learning competitor, while running efficiently on standard CPUs. This work proves that a deep understanding of a problem domain can yield solutions that are not only powerful but also more practical than general-purpose, data-driven approaches.
Il Template Matching è una tecnologia fondamentale nell’automazione industriale, essenziale per task come il controllo qualità e la robotica. Tuttavia, la maggior parte degli algoritmi convenzionali incontra difficoltà nel riconoscere simboli industriali. Questi simboli, infatti, sono disegnati intenzionalmente con forme semplici e colori a tinta unita per favorire la leggibilità per gli esseri umani, finendo però per offrire poche informazioni utili ai metodi tradizionali di Computer Vision. Sebbene il Deep Learning rappresenti una soluzione valida, la sua dipendenza da grandi dataset, i lunghi tempi di addestramento e l’hardware costoso lo rendono impraticabile per ambienti industriali più dinamici. Questa tesi presenta un metodo innovativo e training-free che trasforma questa sfida in un’opportunità. Il nostro sistema expert-driven sfrutta le caratteristiche uniche dei template combinando le loro proprietà cromatiche e geometriche. L’algoritmo utilizza prima le informazioni sul colore per processare efficacemente l’immagine e individuare i potenziali candidati, per poi analizzarne la forma e ottenere un riconoscimento preciso e robusto. I risultati sperimentali dimostrano che il nostro approccio raggiunge valori di accuratezza che superano i metodi concorrenti, incluso un sistema basato su Deep Learning, con il vantaggio di funzionare in modo efficiente su CPU. Questo lavoro dimostra come una profonda comprensione del problema possa portare a soluzioni più efficaci e più pratiche rispetto a più generici approcci data-driven.
Template matching for logos exploiting geometry and color priors
Buccoliero, Francesco
2024/2025
Abstract
Template Matching is a fundamental technology in industrial automation, essential for tasks like quality control and robotic navigation. However, most algorithms struggle to detect common industrial symbols. In fact, these symbols are intentionally designed with simple shapes and plain colors for human readability, which provides little information for conventional Computer Vision methods. While Deep Learning offers a valid solution, its dependence on large datasets, long training, and expensive hardware makes it impractical for many dynamic industrial environments. This thesis presents a novel, training-free method that turns this challenge into an opportunity. Our expert-driven system exploits the unique characteristics of the templates by combining their known color and geometric properties. The algorithm first uses color information to efficiently process the image for potential candidates and then analyzes their shape to achieve precise and robust detection. Experimental results demonstrate that our approach achieves state-of-the-art accuracy, outperforming established methods, including a Deep Learning competitor, while running efficiently on standard CPUs. This work proves that a deep understanding of a problem domain can yield solutions that are not only powerful but also more practical than general-purpose, data-driven approaches.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_10_Buccoliero_Thesis.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
11.42 MB
Formato
Adobe PDF
|
11.42 MB | Adobe PDF | Visualizza/Apri |
|
2025_10_Buccoliero_ExecutiveSummary.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
3.24 MB
Formato
Adobe PDF
|
3.24 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/243280