Utilizzo di reti neurali convoluzionali nella tecnica Structure from Motion

This study proposes an integrated workflow that combines YOLOv11 with Agisoft Metashape to automate photogrammetric pre-processing and the detection of cracks on masonry surfaces with high visual complexity, typical of the architectural and engineering fields. Building on state-of-the art research into the use of YOLO for degradation detection and hybrid detection–segmentation–measurement pipelines, an operational pipeline is developed in which YOLOv11 generates bounding boxes and masks that are subsequently utilised in Metashape via the Python API. The model is trained on a heterogeneous dataset comprising images with annotated cracks and background images. The performance results demonstrate how these technologies can also be utilised in the field of diagnostics within systems employing Structure from Motion algorithms. The detection results are post-processed into highly readable thematic images and reintegrated into the photogrammetric workflow to generate diagnostic textures, transferring information on cracks from 2D to the 3D surface of the structure.