Analyze GA (General Arrangement) Drawings using state-of-the-art deep learning models for automated detection and extraction of tables, nozzles, notes, and views.

• YOLOv8 (nano): For cropping regions of interest (notes, tables, nozzles, views)
• PaddleOCR (OCR + PPStructureV3): For parsing tables and nozzle symbols with high accuracy
• DONUT: Transformer-based model for extracting handwritten/typed notes from images

• GPU: RTX 4070 Ti (12GB VRAM)
• CPU: Intel Core i9-14900KF (24 cores, 32 threads)
• RAM: 64GB

Note: All models currently run on CPU. (Takes ~30 Seconds per image)

{
  "tables": [{ "name": "str", "rows": [["..."]], "bbox": [x, y, w, h] }],
  "nozzles": [{ "bbox": [x, y, w, h], "view": "str", "text": "str" }],
  "notes": { "notes": { "key": "val" }, "bbox": [x, y, w, h] }
}

1. Clone the repository:

   git clone https://www.github.com/acen20/ga-analysis
   cd ga-analysis

   Download det_models/ and ocr_model/ from Google Drive Folder

2. Download YOLOv8 checkpoint files and place them in the det_models/ directory:

   det_models/
   ├── view/
   │   └── best.pt
   ├── section/
   │   └── best.pt
   └── nozzle/
       └── best.pt
   

3. Download DONUT checkpoint files and place them in the ocr_model/ directory:

   ocr_model/
   └── {all extracted files here}
   

4. Build and start the containers:

   docker-compose up --build

5. Access the API endpoint:

   http://localhost:8000/detect
   

Note: First Startup takes time as PaddleOCR downloads OCR models

• The API will be exposed at:
  http://localhost:8000/detect

• Send a POST request with the pdf_file file to the endpoint to get detections and extracted info.

curl -X POST "http://localhost:8000/detect" \
  -F "file=@path/to/your/file.pdf"

1. Limited Dataset Size
   With only 12 documents available for training/testing, the model may not generalize well to variations in document structure, terminology, or formatting.

2. Edge Case – Split Notes Section
   In some cases, the Notes section is split into two parts across the document. This was not detected due to the lack of such variations in the dataset.

   • Potential fix: Oversample relevant cases to train for this pattern. This can be attempted in future iterations.

3. Complex Table Structures
   Some table layouts are too complex for PaddleOCR’s PPStructureV3 to accurately distinguish cell boundaries. A more sophisticated table parsing strategy is required for improved accuracy.