🚀 AnomaVision: Edge-Ready Visual Anomaly Detection

🔥 Lightweight, fast, and production-ready anomaly detection — powered by PaDiM. Deploy anywhere: edge devices, servers, or the cloud.

---

✨ Features

• 📦 Multi-backend exports: PyTorch, ONNX, TorchScript, OpenVINO, TorchRT,INT8 Quantization
• 🎨 Visualizations: anomaly heatmaps, bounding boxes, ROC curves
• 🖥️ Unified Python + CLI workflows
• 🌍 Edge-first design with compact .pth models
• ⚡C++ inference runtime for edge deployment

---

🆚 Why AnomaVision over Anomalib?

• ⚡ 3× faster inference on CPU (MVTec & Visa benchmarks)
• 📦 Smaller models (30 MB vs 40 MB) with lower memory usage
• 🎯 Higher AUROC across most classes on both MVTec AD and Visa datasets
• 🌍 Edge-first design → optimized for ONNX, TorchScript, and OpenVINO
• 🛡️ Production-ready with clean API, CLI, and deployment options

⚡ Speed Comparison (CPU Mode)

Metric	🟢 AnomaVision	🔵 Anomalib	🔺 Difference	🏆 Winner
🧮 Training Time (s)	8.38	13.07	-35.9%	🟢 AV
🚀 Inference FPS	43.41	13.03	+233%	🟢 AV
⏱️ ms / image	23.0	76.7	-70%	🟢 AV

⚡ AnomaVision trains faster, infers over 3× quicker, and processes each image in 70% less time — making it ideal for real-time CPU deployment.

⚡ Speed Comparison (CUDA Mode)

Metric	🟢 AnomaVision	🔵 Anomalib	🔺 Difference	🏆 Winner
🧮 Training Time (s)	8.38	13.07	-35.9%	🟢 AV
🚀 Inference FPS	547.46	355.72	+53.9%	🟢 AV
⏱️ ms / image	1.83	2.81	-35.0%	🟢 AV

⚡ AnomaVision runs over 1.5× faster — ideal for real-time edge inspection.

👉 See detailed results in Benchmarks.

👉 Download: AnomaVision vs Anomalib — A Comprehensive Performance Analysis (PDF)

---

🖥️ C++ Inference with ONNX

AnomaVision isn’t just Python 🚀 — it also provides a C++ implementation for ONNX Runtime + OpenCV.

• 🖼️ Full pipeline: preprocessing → inference → postprocessing → visualization
• 📦 Modular design (Config, Preprocessor, ONNXModel, Postprocessor, Visualizer, App)
• 🌍 Perfect for edge devices and production environments without Python

👉 See full guide: Quickstart — C++ Inference

---

📚 Documentation

📖 Full docs are available in the /docs folder.

• Installation
• Quick Start
• CLI Reference
• API Reference
• Configuration Guide
• Benchmarks
• Troubleshooting & FAQ → Common issues and fixes
• Contributing

---

Installation

Quick Install: using Pyenv and Poetry

git clone https://github.com/DeepKnowledge1/AnomaVision.git
cd AnomaVision
poetry install
poetry shell

From PYPI: PYPI Video

pip install AnomaVision

➡ For detailed installation options, see Installation Guide.

⚡ Quick Example

Train with a config file:

python train.py --config config.yml

## Save:
# Full model → padim_model.pt
# Compact stats-only model → padim_model.pth
# Config snapshot → config.yml

Run detection:

python detect.py --config config.yml

Evaluate performance:

python eval.py --config config.yml

Export to ONNX:

python export.py --config export_config.yml

➡ For more examples, see Quick Start.

---

📊 Benchmarks (Summary)

MVTec AD (15 classes)

• Image AUROC: AV 0.85 ↑ vs AL 0.81
• Pixel AUROC: AV 0.96 ↑ vs AL 0.94
• FPS: AV 43 ↑ vs AL 13

Visa (12 classes)

• Image AUROC: AV 0.81 ↑ vs AL 0.78
• Pixel AUROC: AV 0.96 ↑ vs AL 0.95
• FPS: AV 45 ↑ vs AL 13

📊 Full tables & plots → Benchmarks

---

🤝 Contributing

We welcome contributions!

• See Contributing Guide for high-level steps

---

🙏 Acknowledgments

AnomaVision is built on top of the excellent Anodet repository. We thank the original authors for their contributions to open-source anomaly detection research, which laid the foundation for this work.

---

📜 Citation

If you use AnomaVision in your research, please cite:

@software{anomavision2025,
  title={AnomaVision: Edge-Ready Visual Anomaly Detection},
  author={DeepKnowledge Contributors},
  year={2025},
  url={https://github.com/DeepKnowledge1/AnomaVision},
}

---

💬 Community & Support

• 📢 Discussions
• 🐛 Issues
• 📧 deepp.knowledge@gmail.com

---

👉 Start with Quick Start and build your first anomaly detection pipeline in 5 minutes!