🚀 AnomaVision: Edge-Ready Visual Anomaly Detection

🔥 Production-ready anomaly detection powered by state-of-the-art PaDiM algorithm Deploy anywhere, run everywhere - from edge devices to cloud infrastructure

✨ Supported Export Formats

Format	Status	Use Case	Language Support
PyTorch	✅ Ready	Development & Research	Python
Statistics (.pth)	✅ Ready	Ultra-compact deployment (2-4x smaller)	Python
ONNX	✅ Ready	Cross-platform deployment	Python, C++
TorchScript	✅ Ready	Production Python deployment	Python
OpenVINO	✅ Ready	Intel hardware optimization	Python
TensorRT	🚧 Coming Soon	NVIDIA GPU acceleration	Python

---

✨ What's New (September 2025)

• Slim artifacts (.pth): Save only PaDiM statistics (mean, cov_inv, channel indices, layer indices, backbone) for 2–4× smaller files vs. full .pt checkpoints
• Plug-and-play loading: .pth loads seamlessly through TorchBackend and exporter via lightweight runtime (PadimLite) with same .predict(...) interface
• CPU-first pipeline: Everything works on machines without a GPU. FP16 used only for storage; compute happens in FP32 on CPU
• Export from .pth: ONNX/TorchScript/OpenVINO export now accepts stats-only .pth directly
• Test coverage: New pytest cases validate saving stats, loading via PadimLite, CPU inference, and exporter compatibility

---

✨ Why Choose AnomaVision?

🎯 Unmatched Performance • 🔄 Multi-Format Support • 📦 Production Ready • 🎨 Rich Visualizations • 📏 Flexible Image Dimensions

AnomaVision transforms the cutting-edge PaDiM (Patch Distribution Modeling) algorithm into a production-ready powerhouse for visual anomaly detection. Whether you're detecting manufacturing defects, monitoring infrastructure, or ensuring quality control, AnomaVision delivers enterprise-grade performance with research-level accuracy.

---

✨ Benchmark Results: AnomaVision vs Anomalib (MVTec Bottle, CPU-only)

---

✨ Installation

📋 Prerequisites

• Python: 3.9+
• CUDA: 11.7+ for GPU acceleration
• PyTorch: 2.0+ (automatically installed)

🎯 Method 1: Poetry (Recommended)

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

🎯 Method 2: pip

git clone https://github.com/DeepKnowledge1/AnomaVision.git
cd AnomaVision
pip install -r requirements.txt

✅ Verify Installation

python -c "import anodet; print('🎉 AnomaVision installed successfully!')"

🐳 Docker Support

# Build Docker image (coming soon)
docker build -t anomavision:latest .
docker run --gpus all -v $(pwd):/workspace anomavision:latest

---

✨ Quick Start

🎯 Train Your First Model (2 minutes)

import anodet
import torch
from torch.utils.data import DataLoader

# 📂 Load your "good" training images
dataset = anodet.AnodetDataset(
    "path/to/train/good",
    resize=[256, 192],          # Flexible width/height
    crop_size=[224, 224],       # Final crop size
    normalize=True              # ImageNet normalization
)
dataloader = DataLoader(dataset, batch_size=4)

# 🧠 Initialize PaDiM with optimal settings
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = anodet.Padim(
    backbone='resnet18',           # Fast and accurate
    device=device,
    layer_indices=[0, 1],          # Multi-scale features
    feat_dim=100                   # Optimal feature dimension
)

# 🔥 Train the model (surprisingly fast!)
print("🚀 Training model...")
model.fit(dataloader)

# 💾 Save for production deployment
torch.save(model, "anomaly_detector.pt")
model.save_statistics("compact_model.pth", half=True)  # 4x smaller!
print("✅ Model trained and saved!")

🔍 Detect Anomalies Instantly

# 📊 Load test data and detect anomalies (uses same preprocessing as training)
test_dataset = anodet.AnodetDataset("path/to/test/images")
test_dataloader = DataLoader(test_dataset, batch_size=4)

for batch, images, _, _ in test_dataloader:
    # 🎯 Get anomaly scores and detailed heatmaps
    image_scores, score_maps = model.predict(batch)

    # 🏷️ Classify anomalies (threshold=13 works great for most cases)
    predictions = anodet.classification(image_scores, threshold=13)

    print(f"🔥 Anomaly scores: {image_scores.tolist()}")
    print(f"📋 Predictions: {predictions.tolist()}")
    break

🚀 Export for Production Deployment

# 📦 Export to ONNX for universal deployment
python export.py \
  --model_data_path "./models/" \
  --model "padim_model.pt" \
  --format onnx \
  --opset 17

print("✅ ONNX model ready for deployment!")

---

✨ Real-World Examples

🖥️ Command Line Interface

📚 Train a High-Performance Model

# Using command line arguments
python train.py \
  --dataset_path "data/bottle" \
  --class_name "bottle" \
  --model_data_path "./models/" \
  --backbone resnet18 \
  --batch_size 8 \
  --layer_indices 0 1 2 \
  --feat_dim 200 \
  --resize 256 224 \
  --crop_size 224 224 \
  --normalize

# Or using config file (recommended)
python train.py --config config.yml

Sample config.yml:

# Dataset configuration
dataset_path: "D:/01-DATA"
class_name: "bottle"
resize: [256, 224]        # Width, Height - flexible dimensions!
crop_size: [224, 224]     # Final square crop
normalize: true
norm_mean: [0.485, 0.456, 0.406]
norm_std: [0.229, 0.224, 0.225]

# Model configuration
backbone: "resnet18"
feat_dim: 100
layer_indices: [0, 1]
batch_size: 8

# Output configuration
model_data_path: "./distributions/bottle_exp"
output_model: "padim_model.pt"
run_name: "bottle_experiment"

🔍 Run Lightning-Fast Inference

# Automatically uses training configuration
python detect.py \
  --model_data_path "./distributions/bottle_exp" \
  --model "padim_model.pt" \
  --img_path "data/bottle/test/broken_large" \
  --batch_size 16 \
  --thresh 13 \
  --enable_visualization \
  --save_visualizations

# Multi-format support
python detect.py --model padim_model.pt          # PyTorch
python detect.py --model padim_model.torchscript # TorchScript
python detect.py --model padim_model.onnx        # ONNX Runtime
python detect.py --model padim_model_openvino    # OpenVINO

# Or using config file (recommended)
python train.py --config config.yml

📊 Comprehensive Model Evaluation

# Uses saved configuration automatically
python eval.py \
  --model_data_path "./distributions/bottle_exp" \
  --model "padim_model.pt" \
  --dataset_path "data/mvtec" \
  --class_name "bottle" \
  --batch_size 8

# Or using config file (recommended)
python eval.py --config config.yml

🔄 Export to Multiple Formats

# Export to all formats
python export.py \
  --model_data_path "./distributions/bottle_exp" \
  --model "padim_model.pt" \
  --format all

# Or using config file (recommended)
python export.py --config config.yml

🔄 Universal Model Format Support

from anodet.inference.model.wrapper import ModelWrapper

# 🎯 Automatically detect and load ANY supported format
pytorch_model = ModelWrapper("model.pt", device='cuda')        # PyTorch
onnx_model = ModelWrapper("model.onnx", device='cuda')         # ONNX Runtime
torchscript_model = ModelWrapper("model.torchscript", device='cuda')  # TorchScript
openvino_model = ModelWrapper("model_openvino/model.xml", device='cpu')  # OpenVINO

# 🚀 Unified prediction interface - same API for all formats!
scores, maps = pytorch_model.predict(batch)
scores, maps = onnx_model.predict(batch)

# 🧹 Always clean up resources
pytorch_model.close()
onnx_model.close()

🔧 C++ ONNX Integration

// C++ ONNX Runtime integration example

#include <onnxruntime_cxx_api.h>
#include <iostream>
#include <vector>
#include <string>
#include <chrono>
#include <numeric>
#include <algorithm>

.
.
.
.

---

✨ Configuration Guide

🎯 Training Parameters

Parameter	Description	Default	Range	Pro Tip
backbone	Feature extractor	resnet18	resnet18, wide_resnet50	Use ResNet18 for speed, Wide-ResNet50 for accuracy
layer_indices	ResNet layers	[0]	[0, 1, 2, 3]	[0, 1] gives best speed/accuracy balance
feat_dim	Feature dimensions	50	1-2048	Higher = more accurate but slower
batch_size	Training batch size	2	1-64	Use largest size that fits in memory

📏 Image Processing Parameters

Parameter	Description	Default	Example	Pro Tip
resize	Initial resize	[224, 224]	[256, 192]	Flexible width/height, maintains aspect ratio
crop_size	Final crop size	None	[224, 224]	Square crops often work best for CNN models
normalize	ImageNet normalization	true	true/false	Usually improves performance with pretrained models
norm_mean	RGB mean values	[0.485, 0.456, 0.406]	Custom values	Use ImageNet stats for pretrained backbones
norm_std	RGB std values	[0.229, 0.224, 0.225]	Custom values	Match your training data distribution

🔍 Inference Parameters

Parameter	Description	Default	Range	Pro Tip
thresh	Anomaly threshold	13	1-100	Start with 13, tune based on your data
enable_visualization	Show results	false	true/false	Great for debugging and demos
save_visualizations	Save images	false	true/false	Essential for production monitoring

📄 Configuration File Structure

# =========================
# Dataset / preprocessing (shared by train, detect, eval)
# =========================
dataset_path: "D:/01-DATA"               # Root dataset folder
class_name: "bottle"                     # Class name for MVTec dataset
resize: [224, 224]                       # Resize dimensions [width, height]
crop_size: [224, 224]                    # Final crop size [width, height]
normalize: true                          # Whether to normalize images
norm_mean: [0.485, 0.456, 0.406]         # ImageNet normalization mean
norm_std: [0.229, 0.224, 0.225]          # ImageNet normalization std

# =========================
# Model / training
# =========================
backbone: "resnet18"                     # Backbone CNN architecture
feat_dim: 50                             # Feature dimension size
layer_indices: [0]                       # Which backbone layers to use
model_data_path: "./distributions/exp"   # Path to store model data
output_model: "padim_model.pt"           # Saved model filename
batch_size: 2                            # Training/inference batch size
device: "auto"                           # Device: "cpu", "cuda", or "auto"

# =========================
# Inference (detect.py)
# =========================
img_path: "D:/01-DATA/bottle/test/broken_large"  # Test images path
thresh: 13.0                            # Anomaly detection threshold
enable_visualization: true               # Enable visualizations
save_visualizations: true                # Save visualization results
viz_output_dir: "./visualizations/"      # Visualization output directory

# =========================
# Export (export.py)
# =========================
format: "all"                           # Export format: onnx, torchscript, openvino, all
opset: 17                               # ONNX opset version
dynamic_batch: true                     # Allow dynamic batch size
fp32: false                             # Export precision (false = FP16 for OpenVINO)

---

✨ Complete API Reference

🧠 Core Classes

anodet.Padim - The Heart of AnomaVision

model = anodet.Padim(
    backbone='resnet18',              # 'resnet18' | 'wide_resnet50'
    device=torch.device('cuda'),      # Target device
    layer_indices=[0, 1, 2],          # ResNet layers [0-3]
    feat_dim=100,                     # Feature dimensions (1-2048)
    channel_indices=None              # Optional channel selection
)

🔥 Methods:

• fit(dataloader, extractions=1) - Train on normal images
• predict(batch, gaussian_blur=True) - Detect anomalies
• evaluate(dataloader) - Full evaluation with metrics
• evaluate_memory_efficient(dataloader) - For large datasets
• save_statistics(path, half=False) - Save compact statistics
• load_statistics(path, device, force_fp32=True) - Load statistics

anodet.AnodetDataset - Smart Data Loading with Flexible Sizing

dataset = anodet.AnodetDataset(
    "path/to/images",               # Image directory
    resize=[256, 192],              # Flexible width/height resize
    crop_size=[224, 224],           # Final crop dimensions
    normalize=True,                 # ImageNet normalization
    mean=[0.485, 0.456, 0.406],     # Custom mean values
    std=[0.229, 0.224, 0.225]       # Custom std values
)

# For MVTec format with same flexibility
mvtec_dataset = anodet.MVTecDataset(
    "path/to/mvtec",
    class_name="bottle",
    is_train=True,
    resize=[300, 300],              # Square resize
    crop_size=[224, 224],           # Final crop
    normalize=True
)

ModelWrapper - Universal Model Interface

wrapper = ModelWrapper(
    model_path="model.onnx",        # Any supported format (.pt, .onnx, .torchscript, etc.)
    device='cuda'                   # Target device
)

# 🎯 Unified API for all formats
scores, maps = wrapper.predict(batch)
wrapper.close()  # Always clean up!

🛠️ Utility Functions

# 🏷️ Smart classification with optimal thresholds
predictions = anodet.classification(scores, threshold=15)

# 📊 Comprehensive evaluation metrics
images, targets, masks, scores, maps = model.evaluate(dataloader)

# 🎨 Rich visualization functions
boundary_images = anodet.visualization.framed_boundary_images(images, classifications)
heatmap_images = anodet.visualization.heatmap_images(images, score_maps)
highlighted_images = anodet.visualization.highlighted_images(images, classifications)

⚙️ Configuration Management

from anodet.config import load_config
from anodet.utils import merge_config

# Load configuration from file
config = load_config("config.yml")

# Merge with command line arguments
final_config = merge_config(args, config)

# Image processing with automatic parameter application
dataset = anodet.AnodetDataset(
    image_path,
    resize=config.resize,           # From config: [256, 224]
    crop_size=config.crop_size,     # From config: [224, 224]
    normalize=config.normalize,     # From config: true
    mean=config.norm_mean,          # From config: ImageNet values
    std=config.norm_std             # From config: ImageNet values
)

---

✨ Architecture Overview

AnomaVision/
├── 🧠 anodet/                      # Core AI library
│   ├── 📄 padim.py                 # PaDiM implementation
│   ├── 📄 padim_lite.py            # Lightweight runtime module
│   ├── 📄 feature_extraction.py    # ResNet feature extraction
│   ├── 📄 mahalanobis.py          # Distance computation
│   ├── 📁 datasets/               # Dataset loaders with flexible sizing
│   ├── 📁 visualization/          # Rich visualization tools
│   ├── 📁 inference/              # Multi-format inference engine
│   │   ├── 📄 wrapper.py          # Universal model wrapper
│   │   ├── 📄 modelType.py        # Format detection
│   │   └── 📁 backends/           # Format-specific backends
│   │       ├── 📄 base.py         # Backend interface
│   │       ├── 📄 torch_backend.py    # PyTorch support
│   │       ├── 📄 onnx_backend.py     # ONNX Runtime support
│   │       ├── 📄 torchscript_backend.py # TorchScript support
│   │       ├── 📄 tensorrt_backend.py # TensorRT (coming soon)
│   │       └── 📄 openvino_backend.py # OpenVINO support
│   ├── 📁 config/                 # Configuration management
│   └── 📄 utils.py                # Utility functions
├── 📄 train.py                    # Training script with config support
├── 📄 detect.py                   # Inference script
├── 📄 eval.py                     # Evaluation script
├── 📄 export.py                   # Multi-format export utilities
├── 📄 config.yml                  # Default configuration
└── 📁 notebooks/                  # Interactive examples

---

✨ Contributing

We love contributions! Here's how to make AnomaVision even better:

🚀 Quick Start for Contributors

# 🔥 Fork and clone
git clone https://github.com/yourusername/AnomaVision.git
cd AnomaVision

# 🔧 Setup development environment
poetry install --dev
pre-commit install

# 🌿 Create feature branch
git checkout -b feature/awesome-improvement

# 🔨 Make your changes
# ... code, test, commit ...

# 🚀 Submit pull request
git push origin feature/awesome-improvement

📝 Development Guidelines

• Code Style: Follow PEP 8 with 88-character line limit (Black formatting)
• Type Hints: Add type hints to all new functions and methods
• Docstrings: Use Google-style docstrings for all public functions
• Tests: Add pytest tests for new functionality
• Documentation: Update README and docstrings as needed

🐛 Bug Reports & Feature Requests

• Bug Reports: Use the bug report template
• Feature Requests: Use the feature request template
• Questions: Use GitHub Discussions

---

✨ Support & Community

🤝 Getting Help

1. 📖 Documentation: Check this README and code documentation
2. 🔍 Search Issues: Someone might have had the same question
3. 💬 Discussions: Use GitHub Discussions for questions
4. 🐛 Bug Reports: Create detailed issue reports with examples

👥 Maintainers

• Core Team: @DeepKnowledge1
• Contributors: See CONTRIBUTORS.md

🌟 Recognition

Contributors are recognized in:

• CONTRIBUTORS.md file
• Release notes
• GitHub contributors page

---

---

✨ Roadmap

📅 Q4 2025

• 🚀 TensorRT Backend: NVIDIA GPU acceleration
• 📱 Mobile Export: CoreML and TensorFlow Lite support
• 🔧 C++ API: Native C++ library with Python bindings
• 🎯 AutoML: Automatic hyperparameter optimization

📅 Q1 2026

• 🧠 Transformer Models: Vision Transformer (ViT) backbone support
• 🔄 Online Learning: Continuous model updates
• 📊 MLOps Integration: MLflow, Weights & Biases support
• 🌐 Web Interface: Browser-based inference and visualization

📅 Q2 2026

• 🎥 Video Anomaly Detection: Temporal anomaly detection
• 🔍 Multi-Class Support: Beyond binary anomaly detection
• ⚡ Quantization: INT8 optimization for edge devices
• 🔗 Integration: Kubernetes operators and Helm charts

---

✨ License & Citation

📜 MIT License

AnomaVision is released under the MIT License - see LICENSE for details.

📖 Citation

If AnomaVision helps your research or project, we'd appreciate a citation:

@software{anomavision2025,
  title={AnomaVision: Edge-Ready Visual Anomaly Detection},
  author={DeepKnowledge Contributors},
  year={2025},
  url={https://github.com/DeepKnowledge1/AnomaVision},
  version={2.0.46},
  note={High-performance anomaly detection library optimized for edge deployment}
}

🙏 Acknowledgments

AnomaVision builds upon the excellent work of:

• PaDiM: Original algorithm by Defard et al.
• PyTorch: Deep learning framework
• ONNX: Open Neural Network Exchange
• OpenVINO: Intel's inference optimization toolkit
• Anomalib: Intel's anomaly detection library (for inspiration)

---

✨ Related Projects

• Anodet: anomaly detection

---

✨ Contact & Support

🤝 Community Channels

• 💬 GitHub Discussions: Community Forum
• 🐛 Issues: Bug Reports & Features
• 📧 Email: [email protected]
• 📖 Documentation: Wiki

💼 Enterprise Support

For enterprise deployments, custom integrations, or commercial support:

• 🏢 Enterprise Consulting: Available upon request
• 🎓 Training Workshops: Custom training for your team
• 🔧 Custom Development: Tailored solutions for your use case

---

🚀 Ready to Transform Your Anomaly Detection?

Stop settling for slow, bloated solutions. Experience the future of edge-ready anomaly detection.

---

🏆 Benchmark Results Don't Lie: AnomaVision Wins 10/10 Metrics Deploy fast. Detect better. AnomaVision.

Made with ❤️ for the edge AI community