Enhancing concrete defect segmentation using multimodal data and Siamese Neural Networks
Sandra Pozzer, Gabriel Ramos, Ehsan Rezazadeh Azar, Ahmad Osman, Ahmed El Refai, Fernando López, Clemente Ibarra‐Castanedo, Xavier Maldague
Abstract
This paper proposes an approach for the reliable identification of subsurface damages in thermal images of concrete structures. The work explores how to mitigate false positives in subsurface delamination segmentation using thermal and visible images. The methodology employs a few-shot learning method, specifically the Siamese Neural Network (SNN), to assess the similarity between corresponding multimodal regions. The findings indicate that leveraging similarities between visible and thermal images reduces false positives and improves the segmentation model’s precision by 3.6%, eliminating 351 false positives. These results enhance the reliability of semi-automatic models for detecting subsurface delamination using infrared thermography, benefiting infrastructure maintenance and encouraging the research and development of compact and reliable automation models that integrate civil engineering, nondestructive testing, and artificial intelligence domains.