A framework for synthetic image generation and augmentation for improving automatic sewer pipe defect detection

Sewer pipes are essential infrastructure for discharging wastewater. Regular pipe inspection is necessary to prevent malfunction of sewer systems, which closed-circuit television (CCTV) crawlers commonly used capture images the interior. As manual assessment condition labor-intensive and time-consuming, automated defect detection using computer vision deep learning has been increasingly studied in recent years. However, approaches require large amount annotated data model training. Data collection underground expensive difficult since they inaccessible without use an robot. Meanwhile, ground-truth annotation needs be accurate consistent, requiring massive time expertise. This paper proposes a framework synthetic generation augmentation address shortage problem detection. First, generated by 3D modeling simulation virtual environment. The quality then enhanced style transfer with reference real images. In addition, contrastive module developed further improve process Experiment results show that average precision (AP) improved 2.7% 4.8% respectively after adding style-transferred applying module. When both methods applied, AP boosted 7.7%, from 22.22% 23.92%, indicating effectiveness our proposed approaches. study expected alleviate burden on models