A lightweight cross-scale feature fusion model based on YOLOv8 for defect detection in sewer pipeline

Sewer pipeline defect detection is a critical task for ensuring the normal operation of urban infrastructure. However, the sewer environment often presents challenges such as multi-scale defects, complex backgrounds, lighting changes, and diverse defect morphologies. To address these issues, this paper proposes a lightweight cross-scale feature fusion model based on YOLOv8. First, the C2f module in the backbone network is replaced with the C2f-FAM module to enhance multi-scale feature extraction capabilities. Second, the HS-BiFPN module is adopted to replace the original structure, leveraging cross-layer semantic fusion and feature re-weighting mechanisms to improve the model's ability to distinguish complex backgrounds and diverse defect morphologies. Finally, DySample is introduced to replace traditional sampling operations, enhancing the model's ability to capture details in complex environments. This study uses the Sewer-ML dataset to train and evaluate the model, selecting 1,158 images containing six types of typical defects (CK, PL, SG, SL, TL, ZW), and expanding the dataset to 1,952 images through data augmentation. Experimental results show that compared to the YOLOv8n model, the improved model achieves a 3.8% increase in mAP, while reducing the number of parameters by 35%, floating-point operations by 21%, and model size by 33%. By improving detection accuracy while achieving model lightweighting, the model demonstrates potential for application in pipeline defect detection.