MVTec Anomaly Detection Project

📌 Project Overview

This project aims to implement anomaly detection using the MVTec Anomaly Detection Dataset. It applies various machine learning and deep learning techniques, including ResNet50 feature extraction, KNN-based anomaly scoring, Autoencoder, and synthetic data generation to improve model performance.

📌 Live Demo: MVTec Website

---

🔍 Dataset

The MVTec Anomaly Detection Dataset consists of 15 categories of industrial components and textures. Each category includes:

• Normal images (defect-free)
• Anomalous images (with various real-world defects)

To improve model training and validation, the dataset has been pre-processed, and additional synthetic and augmented data have been generated.
You can download the dataset directly from the official MVTec website.

---

🏗️ Project Structure

---

├── LICENSE
├── README.md          <- The top-level README for developers using this project.
├── data               <- to be dowloaded https://www.mvtec.com/company/research/datasets/mvtec-ad
│
├── models             <- Trained and serialized models, model predictions, or model summaries
│
├── notebooks          <- Jupyter notebooks.
│
├── bibliography       <- Data dictionaries, manuals, links, and all other explanatory materials.
│
├── reports            <- The reports 
│   └── figures        <- Generated graphics and figures to be used in reporting
│
├── environment.yml   <- The yaml file for conda environment: requirements
│
├── src                <- Source code for use in this project.
│   ├── __init__.py    <- Makes src a Python module
│   │
│   ├── features       <- Scripts to turn raw data into features for modeling
│   │
│   ├── models         <- Scripts to train models and then use trained models to make predictions              
│   │
│   ├── streamlit  <- Streamlit presentation
│   │
│   ├── visualization  <- Scripts to create exploratory and results oriented visualizations
│   │   

---

📊 Exploratory Data Analysis & Preprocessing

1. Feature Extraction: Utilizes ResNet50 for deep feature representation.
2. PCA Dimensionality Reduction: PCA retains 95% of variance, which reduces computation time while preserving key data characteristics.
3. Dataset Augmentation: Created 1500 augmented images per category using transformations like flipping, scaling, rotation, and Gaussian blur.

---

🛠️ Machine Learning Models

1. Classical Machine Learning Models + ResNet50

• OC-SVM (One-Class SVM):
  Learns a boundary around the “normal” data in the extracted feature space.

• Isolation Forest:
  Uses a random partitioning process to isolate outliers in the feature representation.

• LOF (Local Outlier Factor):
  Computes local density deviations to flag anomalies based on the feature vectors.

• Elliptic Envelope:
  Assumes a Gaussian distribution and detects points deviating from it within the extracted features.

2. KNN + ResNet50 Feature Extraction

• KNN (k=1) computes anomaly scores by measuring distances in feature space.
• A memory bank of normal samples facilitates efficient lookups.
• Euclidean distance from nearest neighbors determines the anomaly score, with a threshold tuned for optimal F1-score.

3. Autoencoder + ResNet50 Feature Extraction

• An Encoder-Decoder network is trained on normal image feature vectors.
• Reconstruction error (difference between original and reconstructed features) indicates anomalies.
• A threshold based on error distribution balances recall and precision.

---

🔬 Synthetic Data vs. Augmented Data

• Synthetic Data:

  • Created by injecting artificial distortions into normal images.
  • Broadens validation sets when real anomaly samples are limited.

• Augmented Data:

  • Offline augmentation using flips, rotations, scaling, and Gaussian noise.
  • 1500 augmented images per category were pre-generated to improve training diversity.

---

📈 Results & Evaluation

• Primary Metric: AUC-ROC (Area Under the ROC Curve).
• Threshold Optimization: F1-score was used to pick an anomaly detection cutoff.
• Model Comparisons: Assessed via AUC-ROC for each method (OC-SVM, Isolation Forest, LOF, Elliptic Envelope, and KNN with ResNet50 features).
• Visualization: Anomaly score histograms reveal a marked gap between normal and anomalous data.

🚀 How to Run the Project

1. Clone the repository:

   git clone https://github.com/DataScientest-Studio/nov24_bds_mvtec-anomaly-detection.git
   

2. Install dependencies:

   pip install -r requirements.txt
   

3. Explore and train:
   • Navigate to notebooks/
   • Launch Jupyter notebooks to run data exploration and training scripts.

📌 References

• Liu, J., Xie, G., Wang, J., Li, S., Wang, C., Zheng, F., & Jin, Y. (2023). Deep Industrial Image Anomaly Detection: A Survey. Springer Nature. [arXiv:2301.11514].
• Yang, J., Shi, Y., & Qi, Z. (2020). DFR: Deep Feature Reconstruction for Unsupervised Anomaly Segmentation. [arXiv:2012.07122].

• Bühler, J., Fehrenbach, J., Steinmann, L., Nauck, C., & Koulakis, M. (2024). Domain-independent detection of known anomalies. Karlsruhe Institute of Technology (KIT) & preML GmbH. [arXiv:2407.02910].
• Rippel, O., Mertens, P., & Merhof, D. (2020). Modeling the Distribution of Normal Data in Pre-Trained Deep Features for Anomaly Detection. RWTH Aachen University. [arXiv:2005.14140].
• Bergmann, P., Fauser, M., Sattlegger, D., & Steger, C. (2019). MVTec AD – A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection. MVTec Software GmbH. www.mvtec.com.
• Roth, K., Pemula, L., Zepeda, J., Schölkopf, B., Brox, T., & Gehler, P. (2022). Towards Total Recall in Industrial Anomaly Detection. University of Tübingen & Amazon AWS. [arXiv:2106.08265].
• Zheng, Y., Wang, X., Qi, Y., Li, W., & Wu, L. (2022). Benchmarking Unsupervised Anomaly Detection and Localization. University of Chinese Academy of Sciences, SenseTime Research, & Tsinghua University. [arXiv:2205.14852].