Lightweight Technique for Internal Defect Identification of Water Supply Pipelines Based on Convolutional Neural Network

Pressurized internal inspection technology of water supply pipelines has been developed in recent years, enabling a straightforward identification of internal defects. However, these defects are primarily identified by manual methods at present, resulting in heavy workloads and a technique that is prone to errors. For automatic identification on mobile devices, a lightweight technique was proposed in this study by incorporating three key innovations into the original ShuffleNet-V2 structure: (1) hybrid dilated convolutions (HDC) for multiscale feature preservation; (2) efficient channel attention (ECA) for targeted feature enhancement, and (3) H-swish activation function for nonlinear optimization. Building on the proposed lightweight technique, a model named Pipe-ShuffleNet was developed to accurately and efficiently identify internal defects of water supply pipelines, utilizing an image database collected from districts of Shanghai, China between July 2022 and October 2023. To validate the advantages of the developed model, comparative experiments were conducted against six commonly used models: AlexNet, EfficientNet V2, MobileNet V3, ResNet34, GoogLeNet, and VGG16. The investigation demonstrated the advantages of the developed model. On the test dataset, the developed model achieved accuracy of 99.17% and F1 score of 0.9917, while maintaining a low computational overhead with 2.05×106 parameters and 126.72×106 floating point operations per second (FLOPs). Its maximum frames per second (FPSpractical) accounting for the real-world constraints of the computational resources is only 5.68×103. Therefore, the developed model is powerful for deploying on mobile devices to identify internal defects of water supply pipelines accurately in real time.