The phenomenon of tire surface cracking, caused by exposure to degrading agents such as temperature, humidity, and ozone, is of particular interest due to its implications for vehicle safety. To investigate this phenomenon under controlled conditions, a climatic chamber capable of performing accelerated aging tests on tires has been developed at Politecnico di Milano. To monitor damage evolution during the tests, an image acquisition system capable of capturing high-resolution images of specific tire regions was developed. In this work, an upgraded version of the system is presented, designed to accommodate tires with a wider range of dimensions while improving image quality and acquisition repeatability. The acquired images are subsequently subjected to a preprocessing stage including image registration and cropping. For this purpose, a registration algorithm based on the Fast Fourier Transform (FFT) was introduced, demonstrating greater robustness and computational efficiency than the previously adopted Speeded-Up Robust Features (SURF) algorithm. The processed images are then analyzed using convolutional neural networks for crack segmentation. In particular, a comparative study was carried out involving three segmentation architectures: Attention U-Net, Residual Attention U-Net, and DeepLabV3+. The results showed that the Residual-Attention U-Net achieved the best overall segmentation performance. Finally, an uncertainty analysis was performed on two damage-related metrics, namely the Relative Crack Area (RCA) and the Relative Crack Length (RCL). The study showed comparable uncertainty levels among the three architectures, with the Residual-Attention U-Net exhibiting greater robustness in RCA estimation and DeepLabV3+ providing the most robust RCL estimates.
Il fenomeno della criccatura superficiale degli pneumatici, causato dall'esposizione ad agenti degradanti quali temperatura, umidità e ozono, è oggetto di studio per le sue implicazioni sulla sicurezza dei veicoli. Presso il Politecnico di Milano è stata sviluppata una camera climatica in grado di eseguire prove di invecchiamento accelerato sugli pneumatici, consentendo di investigare il fenomeno in condizioni controllate. Per monitorare l'evoluzione del danneggiamento durante i test, è stato realizzato un sistema di acquisizione fotografica capace di acquisire immagini ad alta risoluzione di specifiche aree dello pneumatico. In questo lavoro viene presentata una versione aggiornata del sistema, progettata per ospitare pneumatici di dimensioni più variabili e per migliorare sia la qualità delle immagini sia la ripetibilità degli scatti. Le immagini acquisite vengono successivamente sottoposte a una fase di pre-processamento comprendente allineamento e ritaglio. A tal fine è stato introdotto un algoritmo di registrazione basato sulla Fast Fourier Transform (FFT), che ha dimostrato maggiore robustezza ed efficienza rispetto al precedente approccio basato sull'algoritmo Speeded-Up Robust Features (SURF). Le immagini così processate vengono quindi analizzate mediante reti neurali convoluzionali per il riconoscimento delle cricche. In particolare, è stata condotta un'analisi comparativa tra tre architetture di segmentazione: Attention U-Net, Residual Attention U-Net e DeepLabV3+. I risultati ottenuti hanno evidenziato come la Residual Attention U-Net rappresenti la soluzione più efficace in termini di accuratezza di segmentazione. Infine, è stata eseguita un'analisi dell'incertezza associata a due metriche di danneggiamento, la Relative Crack Area (RCA) e la Relative Crack Length (RCL). Lo studio ha mostrato prestazioni complessivamente comparabili tra le tre architetture, con una maggiore robustezza della Residual-Attention U-Net nella stima della RCA e della DeepLabV3+ nella stima della RCL.
Tire aging: image analysis framework and comparative study of neural networks for crack segmentation
MAINARDI, EDOARDO GIOVANNI
2025/2026
Abstract
The phenomenon of tire surface cracking, caused by exposure to degrading agents such as temperature, humidity, and ozone, is of particular interest due to its implications for vehicle safety. To investigate this phenomenon under controlled conditions, a climatic chamber capable of performing accelerated aging tests on tires has been developed at Politecnico di Milano. To monitor damage evolution during the tests, an image acquisition system capable of capturing high-resolution images of specific tire regions was developed. In this work, an upgraded version of the system is presented, designed to accommodate tires with a wider range of dimensions while improving image quality and acquisition repeatability. The acquired images are subsequently subjected to a preprocessing stage including image registration and cropping. For this purpose, a registration algorithm based on the Fast Fourier Transform (FFT) was introduced, demonstrating greater robustness and computational efficiency than the previously adopted Speeded-Up Robust Features (SURF) algorithm. The processed images are then analyzed using convolutional neural networks for crack segmentation. In particular, a comparative study was carried out involving three segmentation architectures: Attention U-Net, Residual Attention U-Net, and DeepLabV3+. The results showed that the Residual-Attention U-Net achieved the best overall segmentation performance. Finally, an uncertainty analysis was performed on two damage-related metrics, namely the Relative Crack Area (RCA) and the Relative Crack Length (RCL). The study showed comparable uncertainty levels among the three architectures, with the Residual-Attention U-Net exhibiting greater robustness in RCA estimation and DeepLabV3+ providing the most robust RCL estimates.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_Mainardi_Thesis.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
39.15 MB
Formato
Adobe PDF
|
39.15 MB | Adobe PDF | Visualizza/Apri |
|
2026_07_Mainardi_Executive_Summary.pdf
non accessibile
Descrizione: Executive Summary della tesi
Dimensione
688.19 kB
Formato
Adobe PDF
|
688.19 kB | 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/10589/261466