The author uses the full convolutional neural network model for automatic identification of building crack defects. After processing the building crack defect data, the full convolutional neural network model is trained and optimized, and the automatic identification effect of the full convolutional neural network model in this paper is compared with other methods through experiments. In the calculation of building crack defects, image morphology is used to skeletonize the building cracks and calculate the crack length and width. The method is utilized to calculate the physical dimensions of building cracks and their static and dynamic distribution. The average training and validation accuracies of this paper’s model at an initial learning rate of 1e-5 are 95.79% and 95.28%, respectively; the maximum training and validation accuracies are 81.15% and 77.77%, respectively; the maximum training and validation recalls are 80.81% and 80.48%, respectively; and the maximum training and validation F1 values are 84.19% and 79.11%, respectively. Its automatic identification effect is better than other methods. The relative error of maximum crack width between the model prediction results and real results in this paper is 2.1%~25.4%, and the relative error of crack length is 26.48%.The distribution ranges of static crack widths at three locations are between 1.3~3.4mm, 1.02~2.03mm, and 0.42~1.39mm, respectively. The dynamic crack width and area values showed an overall decreasing trend.
