**3. Model-Development Workflow for Crack Detection Based on Faster R-CNN**

In this work we have developed a predictive model for crack detection based on Faster R-CNN architecture described in the previous section. The development of the model consisted of several stages. In the first stage we collected and annotated several images of concrete elements with surface cracks. For model development all annotated images were divided randomly into three sets. The first set was used for training, the second for validation images and third for testing the Faster R-CNN model. Finally, the developed model was deployed for crack detection during experiments described in the next section.
