Lightweight strategy of pipeline detection model based on parameter sharing, pruning and distillation

As a vital component of urban infrastructure, the routine inspection and maintenance of pipelines are essential for ensuring urban safety. However, the high computational demands of existing detection models pose challenges for deployment on resource-limited embedded devices. This study proposes a lightweight optimization strategy for YOLOv8n, which significantly reduces model size while maintaining detection performance. First, a data augmentation technique tailored for pipeline defect detection is introduced to expand the dataset and lower annotation costs. Then, the FasterNet structure incorporating partial convolution (PConv) is employed to modify YOLOv8n’s C2f network, forming the C2f-Light Faster (C2f-LF) structure. Additionally, a lightweight shared parameters detection head (SPDH) is designed, utilizing shared convolution to minimize parameter redundancy while preserving detection accuracy. Furthermore, a channel pruning strategy is implemented to eliminate less significant channels, enhance computational efficiency, and integrate distillation techniques to mitigate accuracy loss from model compression. Experimental results on the real urban pipeline dataset demonstrate that, compared to the baseline YOLOv8n, the model size, parameter count, and floating point operations (FLOPs) are reduced by 86.7%, 89.7%, and 69.1%, respectively. The final model is only 0.8MB. Comparative experiments further validate the proposed method’s effectiveness, offering a highly efficient solution for pipeline detection.