EdgeFlow: Edge-Map Augmented VLM-Based Flowchart Processing for Industrial Requirements Engineering

EdgeFlow is a new VLM-augmented approach that improves flowchart-to-diagram conversion for industrial requirements engineering by incorporating Canny edge detection as a structural prior, achieving significant accuracy gains without requiring model fine-tuning or training data.

EdgeFlow addresses a practical bottleneck in industrial requirements engineering where flowcharts remain trapped as static images rather than machine-readable models. The approach leverages Vision Language Models but augments them with deterministically extracted edge maps to preserve topology—the critical structural relationships that make flowcharts meaningful. This represents pragmatic AI engineering: rather than collecting annotated datasets or retraining models, the researchers identify that VLMs struggle with topology-critical details and inject domain knowledge through preprocessing.

The motivation stems from limitations in applying commodity VLMs directly to specialized tasks. Flowcharts require precise node-to-node connections and pathway accuracy that general vision models optimize poorly. By extracting Canny edges—a classical computer vision technique—as a structural prior, EdgeFlow preserves the topological information that VLMs tend to lose or distort. The 17-point F1 improvement for nodes and 16-point improvement for edges represents meaningful gains in a domain where topology errors cascade through downstream analysis.

The broader significance lies in demonstrating that augmenting modern AI systems with classical domain knowledge remains valuable. For enterprises managing large repositories of legacy flowcharts, this approach offers immediate utility without infrastructure investment. However, the cross-dataset evaluation reveals limitations: EdgeFlow shows no significant improvement on synthetic benchmarks, suggesting the technique is tuned to real-world industrial quirks rather than general flowchart conversion. This highlights that effective AI solutions for industrial processes require domain-grounded validation data and that synthetic benchmarks may inadequately capture real-world complexity.

• →EdgeFlow improves VLM flowchart conversion by 17 percentage points in node accuracy without requiring model retraining
• →Canny edge detection acts as a structural prior that helps preserve topology in flowchart-to-Mermaid conversion
• →The approach is training-free and works with off-the-shelf VLMs, reducing implementation barriers for enterprises
• →Poor performance on synthetic benchmarks indicates that industrial requirements engineering tools need real-world datasets for validation
• →Hybrid classical-modern AI approaches may be more effective than pure end-to-end learning for domain-specific tasks