Lightweight pixel-level semantic segmentation and analysis for sewer defects using deep learning

The underground sewer network is a vital public infrastructure in charge of large-scale wastewater collection and treatment. Complex defects can occur in sewer pipes due to various internal and external factors, which increase the demand for frequent inspection. Previous defect detection research mainly depended on manual inspection, which is tedious, costly, and error-prone. This study suggests an automatic pixel-level sewer defect segmentation framework based on DeepLabV3+, which can recognize the defect’s type, location, geometric information and severity. The impacts of various backbones and pre-processing methods on the model’s performance were carefully evaluated. In addition, four state-of-the-art segmentation models (U-Net, SegNet, PSPNet, and FCN) were compared with the presented model to demonstrate its superiority. The experimental results revealed that the DeepLabV3+ with the Resnet-152 backbone structure efficiently identified ten defect types under challenging conditions. The obtained mean pixel accuracy and mean intersection over union (IoU) were 0.97 and 0.68, respectively. In terms of severity analysis, it was revealed that the framework outputs were consistent with the NASSCO pipeline assessment certification program (PACP). In addition, during the testing process, the proposed frame reduction algorithm only required about 16% of the original time required to process an input video. Finally, with a generated detailed report for an inspection video, the suggested framework can offer a decision-making base for more precise and efficient defect inspection and maintenance.