Crack detection is a highly relevant research field in civil engineering; it involves identifying the presence, shape, and distribution of cracks on materials such as concrete or masonry. Initially, this task was carried out through visual inspections or IoT sensors, which, however, entail high costs and logistical challenges in complex contexts (e.g., towers or bridges). The proposed research addresses this issue by applying Deep Learning techniques for crack detection on masonry surfaces, aiming to find an effective and more economical approach. The experimental phase, structured in two stages, employed two datasets: the first, built from scratch by manually collecting and annotating pixel-wise real images of damaged masonry structures in various areas of Emilia-Romagna (Italy); and the second, a subset of the dataset by Dais et al. ([1]). All images were preprocessed and enriched through data augmentation.

In the first stage, several models were trained on the first dataset, comparing different loss functions commonly used in binary segmentation. The results, expressed through metrics such as Mean Intersection over Union (MIoU) and F1-score (F1), show that Tversky-based loss functions perform better, as they effectively handle class imbalance. In fact, this configuration achieved 0.854 MIoU and 0.874 F1. In the second stage, a cross-dataset training with fine-tuning (FT) was performed. The results not only showed an improvement of about +0.07 in MIoU and +0.08 in F1 after FT but also surpassed the performance of models trained solely on our dataset. These metrics further confirm the validity of these techniques, with the model proposed in this study (MurCrackNet) achieving MDice and F1 values of 0.874. Finally, future developments aim to expand the dataset — including images of different materials such as concrete — to build more robust and generalizable models.