Automatic sewer defect detection and severity quantification based on pixel-level semantic segmentation

Abstract

Underground sewer systems are characterized by their large-scale distributions, high coverage densities and complex defect conditions, which put forward higher requirements for system inspection. Conventional manual inspection is labor-intensive, error-prone and cost inefficient. This paper presents a novel DeepLabv3+-based sewer defect detection and severity quantification method to facilitate automated pixel-level segmentation of sewer defects, from which the defect types, locations, geometric properties and severity levels can be assessed. The effects of different backbone networks on DeepLabv3+ detection accuracy and processing speed were investigated. Three state-of-the-art segmentation methods (SegNet, FCN and U-Net) were further compared to confirm the feasibility of the proposed method. Our results showed that the DeepLabv3+ model, especially with the backbone network Resnet-50, was superior in segmenting multiple types of sewer defects under complex conditions. The obtained PA, mIoU, fwIoU and F1-score were 0.9, 0.53, 0.84 and 0.55, respectively. For individual types of defects, the model worked best in identifying the type of residential walls, followed by disjoints and tree roots. In terms of severity quantification, it was shown that 70% of model predictions were consistent with the ground truths, whereas 30% of predictions likely overestimated the severity levels. The proposed method provides decision-making basis for more accurate and effective sewer inspection and maintenance.