This work aimed to develop an automatic deep learning-based defect detector for sewer networks. A large dataset of 14.7 km of annotated sewer pipe images was created, covering 10 different types of defects and one structural element.
The key highlights and insights are:
Annotating defects in sewer pipes is challenging due to their ambiguous and formless nature, requiring careful guidelines and collaboration among experts.
An EfficientDet-D0 object detector was trained on the dataset, leveraging transfer learning from pre-trained weights on the COCO dataset.
The standard mAP metric was found unsuitable for evaluating the detector's performance on formless defects. Instead, a custom "running meters" metric was developed, which focuses on detecting the presence of defects in pipe sections rather than precise localization.
The final detector achieved 83% detection rate on the test set, with only 0.77% of the missed defects being severe. Expert evaluation confirmed the practical usefulness of the detector's outputs.
Certain defect types, especially those with spatial features, were more challenging for the 2D projection-based detector. Potential improvements include using RGB-D cameras and incorporating expert heuristics as post-processing steps.
Overall, this work demonstrates the application of deep learning-based object detection to an important but underexplored engineering domain, providing practical insights on dataset creation and evaluation for such peculiar "objects" as sewer pipe defects.
翻譯成其他語言
從原文內容
arxiv.org
深入探究