Image matching, a fundamentals task in computer vision, aims to establish correspondences between images for applications like 3D reconstruction and SLAM. Traditional methods rely on keypoints but struggle in low-texture scenes, while line segments, though geometrically informative, face challenges due to imprecise localization and repetitive structures. This paper introduces MGTM (Multi-feature Guided Two-stage Matching), a unified Graph Neural Network (GNN) that jointly matches keypoints and line segments in a two-stage framework. By representing keypoints and line endpoints as nodes in a graph, MGTM leverages self-attention, cross-attention, and feature message passing to integrate visual and geometric cues. This enables mutual disambiguation: keypoints aid line matching in ambiguous regions, and lines guide point matching in textureless areas. A novel line-body preprocessing step connects fragmented line segments, while a coarse-to-fine module refines matches to sub-pixel accuracy.Experiments on ScanNet, DTU, and ETH3D datasets demonstrate MGTM’s superiority over state-of-the-art methods. For homography estimation, MGTM achieves a 13.9° median pose error (16% improvement over SuperGlue) and higher AUC scores. In multi-view stereo reconstruction, it reduces 3D alignment error by 17% compared to point-only methods. Ablation studies validate the contributions of line-body preprocessing, feature message passing, and two-stage matching. MGTM’s data-driven approach outperforms handcrafted geometric strategies, proving the effectiveness of jointly reasoning over points and lines. The code and models are available for further research.