PGC-YOLO: A real-time pipeline defect detection framework with multi-feature optimization and image enhancement

This is the authors' abstract. We don't add key points for this paper.

This study introduces an integrated intelligent inspection system for urban drainage pipelines to address challenges associated with low illumination, complex materials, and diverse defect patterns. An optimized KinD++ enhancement method is introduced to improve brightness by 69.3%, average gradient by 57.9%, and suppress noise and color distortion. The PGC-YOLO algorithm is a lightweight detection framework that uses PKI for multi-level feature fusion, GSConv for computational efficiency, and CARAFE and DIOU to enhance feature reconstruction and localization. PGC-YOLO achieves 98.6% mAP and 98.0% recall across six defect categories, outperforming Faster R-CNN, SSD, and YOLO variants, with strong generalization on the NEU-DET and Underwater Pipes datasets. The model is deployed on a Raspberry Pi and integrated into an autonomous pipeline-robot platform. Field tests confirm stable, accurate detection under low-light conditions, supporting a closed-loop workflow from image acquisition and enhancement to real-time defect recognition.