AI-based crack detection and classification to aid the identification of seismic failure mechanisms of masonry churches

The paper presents an automated investigation of damage to the walls of masonry churches, where cracks are detected and classified by geometric types using AI-based algorithms. It is structured in two parts: crack detection (segmentation) and crack classification. The first part develops a segmentation strategy to detect cracks in images by separating them from the background using an image-based method. A deep learning (U-Net) model, pre-trained on ImageNet (an image database) and fine-tuned on an existing dataset comprising approxi-mately 9,000 images, is initially used. To reduce false positives and some misclassifications in the church environment, the model annotates new image windows added to the training da-taset and iteratively refines segmentation results. The second part prepares a dataset to train an AI model for classifying cracks based on geometric features like linearity and orientation. A weighted random walk method automatically generates synthetic cracks, assigning growth weights to guide the formation of realistic polylines. Initially, one set of real cracks provides the weights, then a method for automatic dataset labeling is investigated. Using the K-means clustering algorithm and rasterized cracks with normalized lengths, polylines with similar ge-ometric features are successfully clustered based on orientation.