[3C4] A post-hoc analysis of a deep learning approach for crack identification in concrete structures
M Sohaib¹, M Hasan² and Z Zheng¹
¹Zhejiang Normal University, China
²Robert Gordon University, UK
Deep learning (DL) has demonstrated promising outcomes in a variety of applications, including the structural integrity assessment of concrete structures. The DL-based structural integrity of the concrete using image data is focused on multiple domains, including defect identification, classification and localisation. However, DL models imply trade-offs between precision and explainability. The DL base models designed for the health assessment of concrete structures are considered to be ‘black box’ models as they lack an explainability factor. To make its predictions comprehensible to humans, the authors employ an explainable artificial intelligence approach. The outputs are converted to human-interpretable language, by fitting the model into a decision tree, and the prediction. Defects are visualised using layer-wise relevance propagation-based approaches. These complementary insights provide further context for the prediction outcomes of the model. The proposed analysis provides the domain expert with confidence and the capacity to explain the prediction of the DL model designed for concrete crack detection.
Keywords: crack detection, decision trees, deep learning, post-hoc analysis.
Keywords: crack detection, decision trees, deep learning, post-hoc analysis.
