Automatic detection of multiple defects in deteriorated concrete sewer pipelines: A method based on StyleGAN and MCFN-YOLO

Concrete sewer pipelines are critical components of urban water and wastewater management systems, which suffer from rupture, deposition, deformation, and misalignment during prolonged service, thereby posing risks to their structural integrity and service life. Conventional closed-circuit television (CCTV)-based manual inspection is inefficient and subject to operator bias. This study aims to address the challenges of low detection accuracy and insufficient data sources in automatic pipeline defect inspection within complex internal pipeline environments as well as under insufficient lighting conditions. A method was proposed for automatic multi-defect detection in deteriorated concrete pipelines based on generative adversarial network data augmentation. This approach integrates StyleGAN3-based data augmentation with a Multi-Scale Context Fusion Network (MCFN)-YOLO detection framework. StyleGAN3 is employed to generate high-fidelity synthetic defect images encompassing diverse defect types, thereby expanding dataset scale and diversity. The proposed MCFN-YOLO incorporates a Multi-Scale Context Fusion Network (MCFN) for hierarchical feature extraction, a Fully Connected PAN (FC-PAN) for dense cross-layer fusion, an Efficient Multi-Scale Attention (EMA) mechanism for feature refinement, and an Adaptive Scale-Aware detection head (ASA-Head) for improved scale adaptability. Experimental results demonstrate that the proposed model achieves 96% mAP and 92% F1-score under complex imaging conditions, outperforming the baseline model (YOLOv8), YOLOv7, YOLOv9, YOLOv12, YOLOv13, SSD, and Faster R-CNN. And its inference speed supports real-time processing of 60 FPS CCTV video streams. The StyleGAN3 synthetic images exhibit high visual realism (FID = 22), enabling significant improvements in generalization. A practical inspection platform based on the proposed method has been further developed, supporting the automatic analysis of images, videos, and real-time data streams, thereby providing an effective solution for the intelligent maintenance of urban sewer infrastructure. • A StyleGAN3-based data augmentation strategy was introduced to effectively address dataset scarcity. • A novel called MCFN-YOLO architecture was proposed. Achieves improvement of 12 % in mAP compared to the baseline model. • The proposed model achieved 96% mAP and 92% F1-score, outperforming comparation models in complex pipeline environments. • MCFN-YOLO demonstrated a recall of 91% on public datasets, confirming the model's generalization capability. • An intelligent defect detection platform was developed, enabling real-time video inspection of concrete sewer pipelines.