Inspection and quality control of composite materials, particularly carbon fiber, are critical in fields such as aerospace, automotive, and renewable energy, where structural defects (delaminations, voids, surface irregularities) can compromise the strength and reliability of components. This study proposes a method for defect classification in carbon fiber composites based on deep learning models that leverage the features obtained from photometric stereo. Photometric stereo, a method that estimates surface normals from images captured under varying lighting conditions, enriches the information on surface topology and shading, thereby exposing subtle defect characteristics. Five different convolutional neural network architectures (Xception, InceptionV3, MobileNetV2, EfficientNet and ConvNeXt) were initially tested on RGB images for both binary and multiclass classification, highlighting the importance of data augmentation in reducing the risk of overfitting and increasing accuracy. Subsequently, the same networks were trained on images obtained from photometric stereo, evaluating how patch overlay can improve fine detail detection and classification accuracy. To further deepen the interpretation of the results and verify that the models were indeed detecting areas containing defects, Class Activation Mapping (CAM) was applied, which provides visual maps of regions crucial for predictions. Finally, to further optimize performance and reduce overfitting, class weighting and label smoothing techniques were used, helping to improve generalization. The results obtained demonstrate how the integration of photometric stereo with deep learning techniques represents a promising solution for automating the inspection and quality control of composite materials. This solution provides a solid foundation for future developments in nondestructive inspection methodologies, with potential industrial applications aimed at ensuring ever-higher quality standards. A possible future research direction is explored in the paper "Segmentation Aware Attention Mechanism for Defect Classification of both Virgin and Recycled Carbon Fiber Fabric", of which I am the second author, and which has been accepted for publication at the Composites International Conference 2025.
L’ispezione e il controllo qualità dei materiali compositi, in particolare della fibra di carbonio, sono fondamentali in ambiti come l’aerospaziale, l’automobilistico e le energie rinnovabili, dove difetti strutturali (delaminazioni, vuoti, irregolarità superficiali) possono compromettere la resistenza e l’affidabilità dei componenti. Questo studio propone un metodo per la classificazione dei difetti nelle matrici di fibre di carbonio, basato su modelli di deep learning che sfruttano le informazioni ottenute tramite fotometria stereo. La fotometria stereo, un metodo che stima le normali della superficie da immagini acquisite in condizioni di illuminazione variabili, arricchisce le informazioni sulla topologia e l'ombreggiatura della superficie, mettendo in luce le caratteristiche più sottili dei difetti. Cinque diverse architetture di reti neurali convoluzionali (Xception, InceptionV3, MobileNetV2, EfficientNet e ConvNeXt) sono state inizialmente testate su immagini RGB per la classificazione binaria e multiclasse, evidenziando l’importanza dell'aumento dei dati nel ridurre l'overfitting e aumentare l’accuratezza. Successivamente, le stesse reti sono state addestrate su immagini ottenute da fotometria stereo, valutando come la sovrapposizione delle patch migliori il rilevamento di dettagli sottili e la precisione nella classificazione. Per approfondire l’interpretazione dei risultati e verificare che i modelli individuassero effettivamente le aree contenenti difetti, è stata applicata la Class Activation Mapping, che fornisce mappe visive delle regioni cruciali per le previsioni. Infine, per ottimizzare ulteriormente le prestazioni e ridurre il più possibile l'overfitting, si è fatto ricorso a tecniche di class weighting e label smoothing, contribuendo a migliorare la generalizzazione. I risultati ottenuti dimostrano come l’integrazione della fotometria stereo con tecniche di deep learning rappresenti una soluzione promettente per l’automatizzazione dell’ispezione e del controllo qualità dei materiali compositi. Questa soluzione fornisce una base solida per sviluppi futuri nelle metodologie di ispezione non distruttiva, con potenziali applicazioni industriali volte a garantire standard qualitativi sempre più elevati. Una possibile direzione di ricerca futura è esplorata nell'articolo “Segmentation Aware Attention Mechanism for Defect Classification of both Virgin and Recycled Carbon Fiber Fabric”, di cui sono il secondo autore, e che è stato accettato per la pubblicazione alla Composites International Conference 2025.
Classificazione dei tipi di difetti nei compositi in fibra di carbonio mediante Apprendimento Profondo attraverso le caratteristiche della fotometria stereo
CARPENTIERI, MATTEO
2024/2025
Abstract
Inspection and quality control of composite materials, particularly carbon fiber, are critical in fields such as aerospace, automotive, and renewable energy, where structural defects (delaminations, voids, surface irregularities) can compromise the strength and reliability of components. This study proposes a method for defect classification in carbon fiber composites based on deep learning models that leverage the features obtained from photometric stereo. Photometric stereo, a method that estimates surface normals from images captured under varying lighting conditions, enriches the information on surface topology and shading, thereby exposing subtle defect characteristics. Five different convolutional neural network architectures (Xception, InceptionV3, MobileNetV2, EfficientNet and ConvNeXt) were initially tested on RGB images for both binary and multiclass classification, highlighting the importance of data augmentation in reducing the risk of overfitting and increasing accuracy. Subsequently, the same networks were trained on images obtained from photometric stereo, evaluating how patch overlay can improve fine detail detection and classification accuracy. To further deepen the interpretation of the results and verify that the models were indeed detecting areas containing defects, Class Activation Mapping (CAM) was applied, which provides visual maps of regions crucial for predictions. Finally, to further optimize performance and reduce overfitting, class weighting and label smoothing techniques were used, helping to improve generalization. The results obtained demonstrate how the integration of photometric stereo with deep learning techniques represents a promising solution for automating the inspection and quality control of composite materials. This solution provides a solid foundation for future developments in nondestructive inspection methodologies, with potential industrial applications aimed at ensuring ever-higher quality standards. A possible future research direction is explored in the paper "Segmentation Aware Attention Mechanism for Defect Classification of both Virgin and Recycled Carbon Fiber Fabric", of which I am the second author, and which has been accepted for publication at the Composites International Conference 2025.File | Dimensione | Formato | |
---|---|---|---|
Carpentieri_Matteo.pdf
embargo fino al 24/03/2026
Dimensione
38.37 MB
Formato
Adobe PDF
|
38.37 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/20.500.12608/83190