EdgeFlow: Edge-Map Augmented VLM for Flowchart Processing

EdgeFlow is a method that enhances Vision Language Models (VLMs) for converting static flowchart images into machine-readable Mermaid code, targeting industrial requirements engineering. By augmenting VLM input with a deterministically extracted Canny edge map as a structural prior, it improves topology-critical detail capture without requiring annotated training data or fine-tuning. Evaluated on the IndusReqFlow dataset from real-world requirements, EdgeFlow achieves node-level F1 improvement of 17.39 percentage points, edge-level F1 improvement of 16.94 percentage points, and path-level F1 improvement of 11.06 percentage points over off-the-shelf VLMs. The approach addresses failures in direct VLM application to flowchart conversion, enabling better support for model-based requirements engineering activities.