Artificial intelligence and multimodal imaging for tunnel inspection - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2022

Artificial intelligence and multimodal imaging for tunnel inspection

Résumé

Early detection of damages in tunnels is a major issue in terms of cost and safety. As here is a wide range of possible damages, varying in size and nature, at times at the limit of detectability, automatic detection based on images is a challenging goal. Cracks are the most common one, but water infiltrations and loss of material, such as exposed irons or spalling, are also present. Recent advances in deep learning are opening new ways to solve this demanding task. In this paper, researchers from the Cerema-ENDSUM team and the SPACETEC Datengewinnung GmbH company assess several state-of-the-art convolutional neural networks (CNN) architectures from the field of semantic segmentation. We propose an experimental study that shows quantitatively the good level of performance that can be obtained by using two different imaging modalities of a damage separately, and qualitatively highlights their complementarity. We finally describe how they can be combined in a CNN architecture.
Fichier non déposé

Dates et versions

hal-03936090 , version 1 (12-01-2023)

Identifiants

  • HAL Id : hal-03936090 , version 1

Citer

François Bock, Arnaud Cereyon, Pierre Charbonnier, Philippe Foucher. Artificial intelligence and multimodal imaging for tunnel inspection. WORLD TUNNEL CONGRESS 2022, ITA-AITES, Sep 2022, Copenhague, Denmark. ⟨hal-03936090⟩

Collections

CEREMA
24 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More