Issue 46

F. Bazzucchi et alii, Frattura ed Integrità Strutturale, 46 (2018) 400-421; DOI: 10.3221/IGF-ESIS.46.37 416 today’s image capturing devices (camera and video-frames) [35]. The fovea of the retina eye is similar to a digital sensor camera: a cluster of retina cones that like pixels are light sensitive and with three recording channels [36]. Digital images are essentially matrix that contains the color information at each pixel by a 2 n -bit byte. n provide the order of the image, for example for n = 3, we have 8-bit images with 256 possible colors given by the combination of the three RGB channels. This means that digital image processing is nothing more than a multidimensional algebra on matrices. Among the many ways that an image can be processed, it is worth to note, for Civil Engineering applications, the filtering, recognizing and detecting operations. To filter an image, matrices are usually multiplied by filters of the same dimension of the image, which at each pixel computes the value of a certain neighborhood of the original image. It exists filters for smoothing, sharpening, averaging or contrasting. Recognizing and detecting are operations performed by convolution between the original image with a smaller matrix, called mask. The mask contains fixed information for the object to detect or trained information for object to recognize. The training of a diagnostic CV algorithm commonly follows two ways: Feature Extraction or Deep Convolution Neural Network (DCNN) [37]. The former one, uses constrained learning over some target features obtained by image processing (i.e. object detection). Its computational burden results to be scarce, but it requires to select a priori where and what to search for. DCNNs instead look for correlation over two datasets, one of whom is used as ground truth [38]. This means the prior knowledge is related to the meaning of the image and not to the driving variables, which instead remains hidden. When DL is involved, the system develops a sort of intuition in recognizing common features as a function of the state variables (dimension of the problem). This operation takes the name of pattern recognition [39]. It is now clear that a system that can recognize and detect damages from image capturing and another one that extract and classify them as a function of common variables could be very useful for large scale diagnosing. In fact, AI with this framework has been successfully used in human medicine in the last years [40]. Some attempt to use NN for structural automated inspection of steel bridges [41] and for general damage detection [42] has been recently proposed. Preliminary results of CV and pattern recognition application to the Fossano bypass road are presented. As described above, the Fossano overpass in one of the 108 bridges that puts together the entire Viaduct. This has been built between 1991 and 1997, and apart from the railway overpass and four short ramps, each bridge has the same structural scheme of the collapsed one. After the failure, the viaduct underwent a series of structural investigations: two time-delayed load tests, tomography, dynamic identifications and Georadar scanning. No decisive results have been extracted yet (due to the problems described above), but they produced a lot of data that can be correlated by pattern recognition to visual inspections. Chemical inspections have evidenced that two particular superficial stains could be related to internal conditions of the prestressing cables. Figure 22 : White stains due to grout drainage (a) ; red-brown stains due to corrosion (b) ; results from classification (c) . The spot in Fig. 22(a) exhibits a with a pale color and a layered aspect. This stain testifies a drainage of the grout by water seepage inside the cables. This obviously implicates inadequate water protection but also suggest that the cables lie in