Deep Learning for Automated Defect Detection in Sewer Systems Using Adaptive Image Pre-Processing

Sewer systems are critical urban infrastructure for safeguarding public health, maintaining environmental integrity, and ensuring sustainable water management. However, their continuous operation is increasingly threatened by ageing assets, rapid urbanisation, and operational constraints that limit inspection frequency. Manual inspection methods including direct man-entry into sewer lines and prolonged review of CCTV survey footage are labourintensive, costly, inconsistent, and unsuitable for timely large-scale monitoring. Recent advances in computer vision and deep learning offer opportunities to address these limitations, yet most research models remain confined to controlled environments, with little attention to reproducibility, defect-specific pre-processing, or deployment in operational settings. This research responds to these gaps by designing, validating, and deploying an automated deep learning defect detection tailored to real-world sewer inspection workflows. A positivist, quantitative methodology guided the study, incorporating two systematic literature reviews to analyse prevailing defect detection approaches and identify key challenges in sewer condition assessment. Historical CCTV survey videos and inspection reports were obtained from multiple Sewer Management Companies (SMCs) in the UK, encompassing 16 defect categories across diverse urban environments. A defect-specific pre-processing strategy was developed, applying tailored enhancement techniques such as adaptive histogram equalisation, contrast stretching, and morphological operations, according to the visual and structural characteristics of each defect type. Cumulatively, 25 different defect-specific pre-processing operation was implemented and applied to the research data. This targeted approach aimed to mitigate issues of low contrast, illumination variation, and feature ambiguity, consequently improving data quality and supporting model robustness. Multiple deep learning architectures, including YOLOv5 through YOLOv11, were comparatively evaluated under consistent hyperparameters, dataset splits, and hardware conditions. Metrics such as mean Average Precision (mAP), FLOPs, number of deep network layers, and latency were used to identify the most effective architecture for multi-class sewer defect detection. The YOLOv9 model was selected for its balance of detection accuracy, computational efficiency, and scalability, which achieves notable performance gains over both earlier YOLO versions and approaches documented in prior studies. Specifically, the optimised YOLOv9 achieved a mAP of 93%, underscoring its robustness and adaptability for diverse field conditions. Validation on unseen municipal datasets confirmed strong generalisation capability, supported by class-balanced augmentation techniques. To ensure real-world applicability, the optimised YOLOv9 model was integrated into a Dockerised FastAPI service deployed on GPU-enabled AWS infrastructure. The system processes both uploaded survey videos and real-time image streams, delivering detection results and detailed defect reports through a user-friendly interface. While deployment results demonstrated operational readiness, the study also identified limitations including small dataset size for defects and performance variation on texture-rich classes. Overall, the research advances the technical and practical readiness of AI-powered sewer defect detection, bridging the gap between academic prototypes and scalable operational tools.