Skip to content

pipeline.md

Latest commit

 

History

History
100 lines (80 loc) · 4.26 KB

pipeline.md

File metadata and controls

100 lines (80 loc) · 4.26 KB

Pipeline: End-to-End P&ID Digitization and Connectivity Analysis

A staged, runnable pipeline with CLI is available in garnet/pid_extractor.py. This pipeline performs comprehensive P&ID analysis with the following stages:

Usage

python garnet/pid_extractor.py \
    --image path/to/pid_image.png \
    --coco path/to/coco_annotations.json \
    --ocr path/to/ocr_results.json \
    --arrow-coco path/to/coco_arrows.json \
    --out output/ \
    --stop-after 11

Pipeline Stages

  1. Stage 1: Input Normalization

    • Load P&ID image (BGR format)
    • Apply normalization (deskew, contrast enhancement)
    • Generate binary mask for pipe detection
    • Output: Normalized image, grayscale, binary mask

  2. Stage 2: OCR Discovery

    • Run selected OCR route (EasyOCR, Gemini, PaddleOCR, or OCRMac)
    • Detect text regions and extract text content
    • Output: OCR results with bounding boxes and text

  3. Stage 3: HITL Review (Manual)

    • Reserved for human-in-the-loop review
    • Major equipment bounding boxes supplied through LabelMe/frontend review
    • Note: This stage is managed outside the automatic CLI flow

  4. Stage 4: Object Detection & Fusion

    • Run SAHI-based object detection (YOLOv8/YOLOv11)
    • Process line number fusion and instrument tag fusion
    • Apply topology-marker routing for junction detection
    • Output: Detected objects with bounding boxes, class labels, and confidence scores

  5. Stage 5: Pipe Mask Generation

    • Create provisional pipe mask from binary image
    • Suppress OCR and object regions to focus on pipe regions
    • Output: Pipe mask image and summary statistics

  6. Stage 5b: Pipe Tracing

    • Trace pipe centerlines from mask using computer vision algorithms
    • Start from detected ports and follow pipe paths
    • Handle inline objects (valves, reducers) by jumping over them
    • Detect terminals (equipment, tags, junctions, sheet edges)
    • Output: Traced pipe paths, junctions, and endpoints

  7. Stage 6: Trace Associations

    • Associate traced paths with ports, inline objects, line numbers, instruments
    • Connect detected objects to the pipe network
    • Fill missing line numbers with simulated-HITL placeholders
    • Output: Object-to-pipe associations and connection data

  8. Stage 7: Geometric Graph Assembly

    • Assemble and normalize the geometric trace graph
    • Label page connectors for multi-document support
    • Run graph quality assurance checks
    • Export initial graph payload (v1)
    • Output: Geometric graph representation (nodes, edges, connectivity)

  9. Stage 8: HITL Review Package

    • Build graph/line-number human-in-the-loop review package
    • Prepare items for validation (junctions, connections, line numbers)
    • Output: Review items ready for frontend inspection

  10. Stage 9: Apply Review Decisions

    • Apply human review decisions to correct the graph
    • Pass through unchanged when no decisions exist
    • Output: Corrected graph based on review feedback

  11. Stage 10: Process Exports

    • Generate line lists, equipment connectivity, inline MTO
    • Create inline observations and instrument index
    • Output: Process exports in JSON/CSV formats

  12. Stage 11: Connection Overlay

    • Render connection-pipeline overlay for visual review
    • Create visual representation of connections on original image
    • Output: Connection pipeline overlay image

Configurable Parameters (PipelineConfig)

  • Image processing: DPI, Canny thresholds, binarization settings
  • Detection: Confidence thresholds per class, image size, overlap ratio
  • Text association: Multiplier for bbox diagonal
  • Graph: Connection radius, port counts, angle separation
  • Valve linking: Directional strategy, edge offset, raycast step
  • Template matching: For valve orientation detection
  • Cleanup: Bridge max distance, angle tolerance
  • Tracing: Branch stub length, merge angle tolerance, opposite angle tolerance
  • Crossing resolution: Center blob radius, threshold, marker match distance

Note

This pipeline is designed to be modular, so each step can be run independently or as part of the full digitization workflow. Use --stop-after N to run only specific stages. The stage numbering intentionally skips Stage 3 in the automatic CLI flow because HITL input is managed outside this runner.