Quantitative Evaluation of Crack Damage Based on U<sup>2</sup>-Net and Morphology

Aiming at the problem of uneven classification of apparent crack data and quantification of damage characteristics of concrete structures, a crack damage assessment method based on U²-Net and morphology was proposed. The Dice loss function and the loss weight were fused to construct the loss function, and the loss function was embedded in transfer learning. The U²-Net network was used to train the data set of apparent cracks of concrete members. The skeleton of the predicted crack map was extracted according to the predicted crack binary image, and finally the cracks were divided into linear cracks and mesh cracks. According to different types of cracks, the corresponding characteristic parameters were extracted by morphology, and the error analysis of the corresponding fracture characteristic parameters was carried out to verify the validity of the fracture characteristic parameters. The results showed that: the average relative errors of linear fracture length, maximum width and average width were －1.34%, 1.08% and 2.64%, respectively, and the average relative errors of mesh fracture area and coverage were 0.05% and 0.07%, respectively. It showed that this method had relatively high detection precision and had certain reference value for crack detection of concrete structures.