AI-driven computer vision-based automated repair activity identification for seismically damaged RC columns

Manual visual inspection is the conventional method for post-earthquake damage assessment, which is unsafe, subjective, and prone to human error. This paper presents an automated rapid and non-contact seismic damage state prediction methodology for reinforced concrete columns using crack image analysis. For surface damage quantification, three features of crack texture complexity including percolation, heterogeneity, and Renyi entropy-based dimensions are measured. Various shallow- and deep-learning-rooted algorithms are trained using a large collected experimental database to develop FEMA P-58-compliant repair activity predictive models. Based on the structural parameters, geometric features, and image-extracted indices, 10 groups of input features are defined. For the overfitting assessment and generalizability evaluation of models, five-fold cross-validations are conducted. Among shallow learning-based algorithms, CatBoost algorithm performs best for the scenarios that rely on vision-derived intricacy indices. Using the deep learning-based multilayer perceptron model as a feedforward artificial neural network, 92 % accuracy is achieved for the testing dataset.