Automatic Pipeline Defect Detection Based on A Sewer Robot and Instance Segmentation

Abstract

Automatic defect detection within sewer pipelines is crucial for their normal working. This paper introduces an improved network architecture and a dataset constructed for sewer defect detection through instance segmentation. Our architecture integrates YOLO (You Only Look Once)-based feature extraction layers augmented with ResSPP modules alongside YOLACT (You Only Look At CoefficienTs)-based segmentation mask generators. The proposed dataset comprises 3,958 images annotated by professional sewer inspectors, containing 16 classes derived from the Chinese national standard. The proposed network is optimized to strike a meticulous balance between detection accuracy and inference speed, outperforming existing real-time instance segmentation networks when testing on our dataset. Ablation studies and comparative analyses verified the effectiveness of the proposed architectural components. Experimental results demonstrated that our network obtained a mAP@50 of 86% and a mAP@50:95 of 59.8% on the validation set while maintaining an expeditious inference time of 5.2 milliseconds—fulfilling the demands of sewer defect detection.