PSI - Issue 44

Carpanese Pietro et al. / Procedia Structural Integrity 44 (2023) 1980–1987 Carpanese Pietro et al./ Structural Integrity Procedia 00 (2022) 000–000

1983

4

A Convolutional Neural Network is a class of neural networks specialized for processing data with a grid-like topology, such as images. A CNN is structured so that each layers can recognize simpler patterns first (such as lines and curves) and more complex patterns later (faces and objects). A typical CNN is composed of convolutional layers, pooling layers, and fully connected layers. The convolutional layer performs a dot product between two matrices, one representing the set of learnable parameters, also called the kernel, and the other representing a portion of the image. The pooling layer replaces the network’s output at specific locations by deriving summary statistics of nearby outputs, thus reducing the spatial size of the representation. Lastly, the fully connected (FC) layer maps the representation between the input and the output by converting the image into a column vector. In the following paragraphs, the dataset on which the CNNs are trained is described, as well as the structure of the neural networks. 3.1. Dataset with labeled images of Italian residential buildings In order to train a Convolutional Neural Network, the first step is to create a consistent database of images. For this specific work, many images of Italian residential buildings were collected and labeled according to the different features for which the CNNs needed to be trained: height, material, and construction period. To do so, the following sources have been used. Firstly, the access to the CARTIS web portal was leveraged. This platform collects all the data gathered within CARTIS, a project carried out by the ReLUIS consortium and funded by the Italian Department of Civil Protection. CARTIS proposes a procedure for carrying out municipal and regional inventories based on the typological-structural characterization of buildings (Zuccaro et al. 2015). In the CARTIS web portal, it is possible to download the forms of the buildings already surveyed by the research units involved in this project. These forms often include a picture of the building, as well as information about the number of stories, the material, and the construction period. Another source from which valuable images of buildings were collected is Da.D.O. (Dolce et al. 2019). This platform includes forms that contain data on seismic damage detected on buildings after the main Italian seismic events. Each building is labeled with their main characteristics, which include those needed to create the dataset. Furthermore, each form reports the building coordinates (latitude and longitude), which makes it possible to automatically retrieve the image of the building, for example, through Google Street View. In order to avoid collecting images of buildings that had suffered severe damage or even partial/total collapses, a filter was applied to download only the forms relating to buildings that had not suffered damage, i.e., with damage index D0. The last source from which other labeled pictures of Italian residential buildings were retrieved was an extensive direct survey that was carried out in 2018-2019 by the University of Padova involving the entire Municipality of Pordenone, in the Region of Friuli-Venezia Giulia, North-East of Italy (Vettore et al. 2020). This survey allowed the collection of many labeled pictures of different kinds of buildings belonging to that municipality.

Table 1. Splitting of the images into training, evaluation and validation for the 3 features. Feature Labels Training (70%) Validation (15%) Evaluation (15%)

Total (100%)

Low-Rise Mid-Rise Masonry Pre-1919 1919-1945 1946-1960 1961-1980 Post-1980 R.C.

4,004 2,998 4,906 2,094 1,621

856 643 449 348 109 440 391 211

856 643 449 348 109 440 391 211

Height

10,000

1,051

1,051

Material

10,000

510

Construction period

2,056 1,826

10,000

989

In total, 10,000 pictures of buildings were collected to build the final dataset. Not all the pictures were used to train the CNNs: some of them were kept for validation and evaluation. The same splitting (i.e., 70% to the training set, 15% to the evaluation set, and 15% to the validation set) was implemented for the three different features on which the CNNs were trained. In Table 1, it is possible to see the number of images belonging to each folder.