DPANet: curve-like structure segmentation based on dual-path attention network

Curve-like structure segmentation has wide applications in computer vision, medical image analysis, and industrial fields. However, achieving complete segmentation of curve-like structures remains challenging due to complex background interference. Methods based on global features tend to ignore local details, while methods based on local features often fail to maintain topological correctness, resulting in fragmented segmentation results. In this work, we propose DPANet, a simple and powerful framework for segmenting curve-like structures. We design a Global Multi-branch Attention (GMBA) module that captures cross-dimensional interactive information to enhance segmentation accuracy in complex scenes. We develop a Local Multi-layer Convolution (LMLC) module that effectively preserves structural continuity through local detail extraction. Finally, we develop a Differential Feature Enhancement (DFE) module for compensating for the loss of multi-scale feature information, refining intermediate feature representations, and enhancing the network's ability to capture fine-grained details and structural boundaries. We conducted qualitative and quantitative experiments on five datasets (Yarn Hairiness, DRIVE, XCAD, CrackTree200, and Crack500). The experimental results demonstrate that DPANet achieves superior performance in the segmentation task of complex curve-like structures, and effectively addresses the continuity issues in curve-like structures segmentation.