Intelligent identification and classification of sewer pipeline network defects based on improved RegNetY network

The operation of a sewer pipeline network affects the urban drainage effect, and identifying the defects of the pipe network is very significant to accurately evaluate the condition of the pipeline network. Closed-circuit television detection has been widely used in urban sewer pipeline network inspection. However, manual discrimination requires professionals to have enough knowledge reserves, which is subjective, laborious, and time-consuming. It is necessary to realize automation and intelligence of defect classification of the sewer pipeline network. An intelligent identification model for sewer pipeline network defects is constructed based on an improved RegNetY network and gradient-weighted class activation mapping (Grad-CAM). The defects are classified and visualized. First, data augmentation is used to deal with data imbalance. Then, LeakyReLU is used as the activation function to improve the RegNetY network. Finally, the training features of the improved RegNetY network are visualized using Grad-CAM. The engineering application results show that the improved RegNetY is more competitive than other types of deep neural networks in defect classification, which can improve the inspection efficiency of engineers. The Grad-CAM can visualize the specific location of defects, which is beneficial to understanding the defect information in images intuitively.