VisAdj Learns Graph Adjacency from Node-Link Images.

Extracting structured graph information from visual representations, such as node-link diagrams, is a fundamental challenge. Current approaches often rely on fixed, heuristic methods for identifying potential connections between nodes, which limits their ability to capture complex relationships among edges. This new research introduces VisAdj, a novel framework designed to overcome these limitations through a topology-aware approach to adjacency prediction. VisAdj employs an attention-sparse neighbor sampler to intelligently select a high-recall set of candidate node pairs. This adaptive selection process is then followed by a joint edge inference mechanism utilizing a line-graph transformer. This transformer treats candidate edges as tokens, allowing it to explicitly model and understand the dependencies between incident edges, leading to a more accurate reconstruction of the underlying graph structure. Extensive testing on diverse datasets, including synthetic graphs, real-world road networks, and medical vessel images, demonstrates that VisAdj consistently and significantly outperforms previous state-of-the-art methods. This indicates a substantial advancement in the ability to automatically derive graph structures from visual data.