LitePipeNet: Research on a Lightweight and Efficient Segmentation Model for UAV Pipeline Inspection in Mining Areas

In response to the challenges of UAV-based pipeline inspection in mining areas—including difficulties in small object detection, limited adaptability to complex environments, and constraints on efficient edge deployment—this paper proposes LitePipeNet, a lightweight segmentation model tailored for edge computing. To support model development, we construct the MinePipe-UAV dataset, containing 7,960 high-resolution images with typical inspection targets such as pipelines, corrosion, and leakage. LitePipeNet integrates Multi-Path Weight Convolution (MPWConv) for enhanced perception, a Lightweight Bidirectional Feature Pyramid Network (LBiFPN) for efficient feature fusion, and a Multi-Level Decoupled Segmentation Head (MLHead) to optimize segmentation. This architecture significantly improves small-object recognition and environmental adaptability. Experiments show LitePipeNet achieves a Mask mAP50 of 89.74% on MinePipe-UAV, outperforming mainstream models (YOLOv8n, DeepLabV3+) by 3–6% in segmentation accuracy. With only 2.9M parameters and 10.5 GFLOPs, it reaches real-time inference at 28 FPS on the RKNN 3588 platform, making it well-suited for industrial edge deployment. Extensive comparisons and ablation studies confirm its robustness and generalization. The main contributions are: 1) construction of the MinePipe-UAV dataset for UAV-based pipeline inspection; 2) design of the compact, high-performance LitePipeNet model; 3) systematic validation of our approach’s effectiveness and scalability in industrial visual edge applications.