German-Layout

Kompaktes Document Layout Detection Model für deutsche Dokumente

Features

• Kompakt: 6 MB PyTorch / 12 MB ONNX
• Schnell: <100ms Inference, 453 FPS mit TensorRT
• Präzise: 99.5% mAP50 auf Validierungsdaten
• 10 Klassen: text_block, heading, table, list, image, footer, header, signature, stamp, logo
• Multi-Format: PyTorch, ONNX, OpenVINO, TensorRT

Quick Start

Installation

# Repository klonen
git clone https://github.com/Keyvanhardani/German-Layout.git
cd german-layout

# Virtual Environment erstellen
python -m venv venv
source venv/bin/activate  # Linux/Mac
# oder: venv\Scripts\activate  # Windows

# Dependencies installieren
pip install ultralytics torch opencv-python pillow onnx fastapi uvicorn

Inference

# Einzelnes Bild
python src/inference.py document.png --model models/german_layout_yolov8n.pt --output result.png

# Verzeichnis
python src/inference.py documents/ --model models/german_layout_yolov8n.pt --output results/

REST API

# Server starten
uvicorn src.api:app --host 0.0.0.0 --port 8000

# Health Check
curl http://localhost:8000/health

# Layout Detection
curl -X POST http://localhost:8000/detect -F "file=@document.png"

# Mit Visualisierung
curl -X POST http://localhost:8000/detect/visualize -F "file=@document.png" -o result.png

Python API

from ultralytics import YOLO

# Model laden
model = YOLO("models/german_layout_yolov8n.pt")

# Inference
results = model("document.png")

# Ergebnisse
for r in results:
    for box in r.boxes:
        cls = int(box.cls[0])
        conf = float(box.conf[0])
        x1, y1, x2, y2 = box.xyxy[0].tolist()
        print(f"Class: {model.names[cls]}, Confidence: {conf:.2f}")

Projekt-Struktur

german-layout/
├── config/
│   └── dataset.yaml         # Dataset-Konfiguration
├── data/
│   ├── train/               # Trainingsdaten
│   └── val/                 # Validierungsdaten
├── models/
│   ├── german_layout_yolov8n.pt      # PyTorch
│   ├── german_layout_yolov8n.onnx    # ONNX
│   ├── german_layout_yolov8n_openvino_model/  # OpenVINO
│   └── german_layout_yolov8n.engine  # TensorRT
├── src/
│   ├── api.py               # FastAPI REST API
│   ├── train.py             # Training Script
│   ├── evaluate.py          # Evaluation
│   ├── inference.py         # Inference Script
│   ├── export_onnx.py       # ONNX Export
│   ├── export_openvino.py   # OpenVINO Export
│   ├── export_tensorrt.py   # TensorRT Export
│   └── generate_synthetic_data.py
├── MODEL_CARD.md
└── README.md

Training

# Synthetische Daten generieren
python src/generate_synthetic_data.py --train 500 --val 100

# Training starten
python src/train.py --epochs 50 --batch 16 --device 0

# Evaluation
python src/evaluate.py --model models/german_layout_yolov8n.pt

Export

# ONNX
python src/export_onnx.py --model models/german_layout_yolov8n.pt

# OpenVINO
python src/export_openvino.py --model models/german_layout_yolov8n.pt

# TensorRT (nur NVIDIA)
python src/export_tensorrt.py --model models/german_layout_yolov8n.pt --fp16

API Endpoints

Endpoint	Methode	Beschreibung
/health	GET	Health-Check
/classes	GET	Verfügbare Klassen
/detect	POST	Layout-Erkennung (JSON)
/detect/visualize	POST	Erkennung mit Visualisierung
/detect/base64	POST	Base64-Input
/detect/batch	POST	Batch-Verarbeitung

Performance

Hardware	Inference Time	FPS
RTX 4000 SFF Ada (PyTorch)	~12ms	83
RTX 4000 SFF Ada (TensorRT FP16)	2.2ms	453
Intel CPU (OpenVINO)	~50ms	20

Lizenz

Apache License 2.0

Credits

• YOLOv8: Ultralytics
• System: AEGIS Supervisor
• Architect: Keyvan.ai

---

Erstellt von AEGIS Supervisor - 2025-12-30