metadata

license: apache-2.0
datasets:
  - mvtec-ad
metrics:
  - auroc
  - f1
pipeline_tag: image-segmentation
tags:
  - anomaly-detection
  - industrial-inspection
  - mvtec-ad
  - deep-learning
  - openvino
  - quality-control
library_name: openvino

Model Card for MetalPart-Anomaly-Detector

This model detects anomalies in metal parts during production processes. It uses Deep Learning and OpenVINO Runtime for high-accuracy anomaly detection, providing heatmaps and segmentation masks for visualizing defects like scratches or deformations.

Model Details

Model Description

Developed by: Keyvan Hardani
Shared by: GitHub Repository
Model type: Image segmentation and anomaly detection
License: Apache 2.0
Finetuned from model: None

Model Sources

Repository: GitHub Link
Demo: Hugging Face Demo Link

Uses

Direct Use

This model is directly usable for:

Quality Control: Ensuring defect-free metal parts in production.
Predictive Maintenance: Early detection of anomalies to avoid major breakdowns.
Automated Inspection: Enhancing efficiency in industrial workflows.

Out-of-Scope Use

This model is not suited for non-industrial materials or environments with highly unstructured data.

Bias, Risks, and Limitations

Limitations

Requires high-quality input images with consistent lighting for optimal results.
Performance may vary depending on the dataset used.

Recommendations

Users should test the model with a subset of their own data before large-scale deployment.

How to Get Started with the Model

To use this model:

Download the pre-trained weights (model.xml, model.bin, and metadata.json) from the repository.
Place the model files in the appropriate directory, as described in the GitHub README.

Training Details

Training Data

Dataset Used: MVTec AD (metal parts subset)
Preprocessing: Normalization and resizing to model-specific input dimensions.

Training Procedure

Framework: OpenVINO Runtime
Loss Function: Cross-Entropy Loss
Optimizer: Adam

Evaluation

Metrics

AUROC: Measures the model's ability to distinguish between anomalous and normal parts.
F1 Score: Assesses the balance between precision and recall.

Results

Image AUROC: 0.95
Image F1 Score: 0.94
Pixel AUROC: 0.96
Pixel F1 Score: 0.71

Environmental Impact

Hardware Type: GPU-based training and inference (NVIDIA RTX 4080)
Hours used: Approx. 10 hours
Carbon Emitted: [Estimate pending]

Citation

If you use this model, please cite it as:

@misc {keyvan_hardani_2024, author = { {Keyvan Hardani} }, title = { AnomalyDetection-MVTech-Metal (Revision b326b4e) }, year = 2024, url = { https://huggingface.co/Keyven/AnomalyDetection-MVTech-Metal }, doi = { 10.57967/hf/3678 }, publisher = { Hugging Face } }

Model Card Authors

Keyvan Hardani

Contact

For questions or support, please reach out via GitHub Issues