ISO 5807 Flowcharts Overview

• ISO 5807 Flowcharts are a standardized graphical system that encodes procedural steps using defined symbols and explicit connectivity rules.
• They are applied in maintenance, industrial, and business settings to ensure unambiguous communication through symbols such as rectangles, diamonds, and arrowheads.
• Recent research leverages machine learning and computer vision to automate symbol detection and edge reconstruction, achieving high precision and recall in extracting process flows.

ISO 5807 Flowcharts represent a standardized graphical convention for encoding procedural knowledge in maintenance, industrial, and business process documentation. The ISO 5807:1985 standard prescribes a controlled vocabulary of diagrammatic symbols and explicit connectivity semantics to facilitate clear, unambiguous communication of stepwise procedures. Recent work applies computer vision and LLMs to automate the extraction and formalization of these graphical representations, enabling queryable procedural knowledge essential for modern operator support and asset management systems (Avalle et al., 8 Apr 2026).

1. ISO 5807 Symbol Taxonomy and Semantics

The ISO 5807 standard designates canonical geometric forms for flowchart elements, each assigned distinct operational semantics. In contemporary automated analysis (e.g., FlowExtract), six element classes receive explicit recognition:

Symbol	Geometric Form	Semantics
Process	Axis-aligned rectangle	Single procedural step or action
Decision	Rhombus (diamond)	Binary test or branching logic
Document	Rectangle with sinuous (wavy) bottom edge	External document or reference
Terminator	Capsule shape (rectangle with semicircular ends)	Start/end marker of a procedure
Connector	Circle	On-page/cross-page jump reference
Arrowhead	Small triangular/trapezoidal tip	Directional cue (not a node)

All geometric constraints are managed implicitly by bounding-box detection; the standard specifies general form (e.g., alignment, wavy edges, capsule ends) but omits aspect ratio or size equations. Arrowheads, as prescribed by ISO 5807, strictly indicate flow direction, always pointing from source to target node.

2. Connectivity and Directionality Conventions

ISO 5807 establishes that directed single-headed arrows encode procedural flow. The arrowhead's orientation is intrinsically meaningful: the pointed end denotes the successor (target) node, the blunt end the predecessor (source). Only single-headed arrows (triangular tips) appear in compliant flowcharts for the use case addressed in FlowExtract. There is no ISO 5807 provision for multiple arrowhead styles or alternate vector formulas in basic diagrams. The entire graph topology emerges from the detected arrowheads and their linking lines, as opposed to textual relationship cues.

3. Automated Symbol Detection: Methods and Configurations

Automated processing of ISO 5807 flowcharts employs high-performance object detection pipelines. FlowExtract utilizes the YOLOv8s detector, pre-trained on COCO but fine-tuned for domain-aligned recognition of the six canonical symbol classes. Training incorporates 25 annotated images (823 symbol instances) from Dutch maintenance flowcharts, with severe class imbalance mitigated by mosaic augmentation (mixing quadrants from four images per training sample) and brightness jitter (HSV ±30%) to simulate diverse scan conditions.

Critically, no rotations or flips are applied during augmentation to preserve the natural orientation of arrowheads and symbols. The detection pipeline operates at 640×640 px resolution, with cosine-annealed learning rate scheduling (0.01 initial value, 250 epochs), and explicit batch size management (batch size 8). Precise type classification and bounding localization underpin accurate downstream edge extraction.

4. Edge Detection and Connectivity Reconstruction Algorithm

Connectivity recovery is anchored on explicit arrowhead detection. The pipeline proceeds as follows:

1. For each detected arrowhead, compute its centroid and determine the two boundary extremes along the major axis.
2. The extreme point nearest any node centroid is designated the arrow tip; the other is the base.
3. A direction unit vector $$\vec{v} = \frac{C_{tip} - C_{base}}{\|C_{tip} - C_{base}\|}$$ underlies orientation, although not written explicitly in the paper.
4. From the arrow base, line tracing is performed via the Probabilistic Hough Transform on a binarized diagram image. Graph traversal propagates through line segment endpoints until a node boundary is identified as the source.
5. The JSON output encodes the adjacency list, not an explicit matrix; for documentation purposes, the adjacency relation may be represented as $A_{ij} = 1$ if node $i$ links to node $j$ via a directed edge, $A_{ij} = 0$ otherwise.

Edge labels, such as decision branch names (“ja”/“nee”), are assigned based on OCR-proximal text to the arrow’s midpoint.

5. Evaluation Metrics and Comparative Performance

Empirical evaluation on a held-out test set (7 images, 219 annotated elements) demonstrates the efficacy and limitations of the approach:

• Node detection across the five flowchart symbol classes yields 98.4% precision, 99.2% recall, and 98.8% F1 score; type classification achieves 97.6% accuracy.
• Arrowhead detection accuracy, critical to successful edge extraction, attains 73.1%.
• Text extraction (EasyOCR) on both node interiors and standalone labels reaches 99.2% character accuracy.
• Edge reconstruction achieves 85.5% precision, 54.6% recall, and 66.7% F1; edge-label detection sits at 73.8% accuracy.

In direct comparison to Vision-LLM baselines, which achieve at most Node F1=0.34 and Edge F1=0.11, FlowExtract reports a six-fold improvement in edge extraction precision and recall. The precision-first orientation prioritizes skeletonization accuracy, supporting reliable graph extraction for downstream human-in-the-loop validation (Avalle et al., 8 Apr 2026).

6. Practical Considerations and Limitations

FlowExtract’s pipeline design, favoring precision in both node and edge proposals, is optimized for practical industrial deployments in scenarios where procedural correctness is paramount. Conservative edge generation reduces human verification burden by shifting the residue of uncertainty toward omission (missing edges) rather than commission (incorrect edges), aligning with ISO 5807’s emphasis on unambiguous symbol semantics. This approach facilitates rapid integration into asset-management and operator-support systems.

Primary limitations are tied to the reliability of arrowhead detection under occlusion or when rendered at very small sizes, which constrains recall, and the absence of explicit cross-document connector resolution. These represent salient directions for subsequent refinement, especially as procedural complexity increases or documentation artifacts span multiple pages.

7. Standards Alignment and Future Directions

FlowExtract demonstrates faithful adherence to ISO 5807 geometry and semantics by learning symbol morphology directly from annotated examples and encoding diagram topology exclusively via arrowhead-oriented connectivity heuristics. Open domains for future work include improved handling of small/occluded arrowheads, formal cross-document node linking, and generalization to other ISO flowcharting conventions. As procedural knowledge digitization efforts scale across industrial domains, the robustness and standard conformity of the flow extraction pipeline provide a template for standards-aligned knowledge base construction and automated process validation (Avalle et al., 8 Apr 2026).