P&IDSeg-Net:一种面向管道仪器图的轻量化分割网络

To enhance the performance of intelligent segmentation for piping and instrumentation diagrams (P&JDs), a lightweight segmentation network named P&TDSeg-Net is proposed. First, a channel-halved non-local attention module was designed lo establish global dependencies in deep feature space, addressing ihe structural confusion between pipelines and instruments. Second, feature addition fusion was employed instead of channel concalenation for skip connections, furlher reducing the parameter count and adapting to small-scale dalasets. Then, based on the low-grayness characteristic of pipeline regions in P&TDs, a visual prompt-inspired grayscale mask weighted loss was constructed lo focus the nelwork on key foreground areas. Finally, the P&TD_Pipe dataset was collected and manually annotated, and extensive experiments were conducted on this dataset. The results show thai ihe proposed P&-IDSeg-Net requires only 2. 181 × 10⁷ parameters, demonstrating significant lightweight advantages, while achieving a mean interseclion over union of 77. 40%. Compared with mainstream segmentation methods, it achieves superior performance.