Pixel-level multicategory detection of visible seismic damage of reinforced concrete components

The detection of visible damage (i.e., cracking, concrete spalling and crushing, reinforcement exposure, buckling and fracture) plays a key role in postearthquake safety assessment of reinforced concrete (RC) building structures. In this study, a novel approach based on computer-vision techniques was developed for pixel-level multicategory detection of visible seismic damage of RC components. A semantic segmentation database was constructed from test photos of RC structural components. Series of datasets were generated from the constructed database by applying image transformations and data-balancing techniques at the sample and pixel levels. A deep convolutional network architecture was designed for pixel-level detection of visible damage. Two sets of parameters were optimized separately, one to detect cracks and the other to detect all other types of damage. A postprocessing technique for crack detection was developed to refine crack boundaries, and thus improve the accuracy of crack characterization. Finally, the proposed vision-based approach was applied to test photos of a beam-to-wall joint specimen. The results demonstrate the accuracy of the vision-based approach to detect damage, and its high potential to estimate seismic damage states of RC components.