cnn-autoencoder

Industrial Sense AI: Unsupervised Visual Inspection & Anomaly Detection

An end-to-end unsupervised visual inspection framework built with PyTorch and Gradio. This system utilizes a Convolutional Autoencoder with a compressed bottleneck vector to detect, locate, and measure manufacturing defects on high-resolution industrial components (e.g., MVTec AD dataset).

---

Key Features

• Convolutional Autoencoder Architecture: Features 4 convolutional encoder stages down to a 128-dimensional bottleneck, paired with 4 transpose-convolutional decoder stages.
• Unsupervised Anomaly Detection: Learns standard feature representations exclusively from defect-free training images. Anomalous regions produce elevated reconstruction errors.
• Pixel-Level Heatmap Overlay: Computes L1 reconstruction distance map, applies Gaussian smoothing, and overlays colorized heatmaps (JET colormap) on original input images.
• ECDF Error Diagnostics: Generates Empirical Cumulative Distribution Function (ECDF) curves and diagnostic charts comparing normal vs. defective pixel error distributions.
• Interactive Gradio Interface: Standalone web UI allowing users to upload inspection images, select dataset samples, adjust anomaly decision thresholds dynamically, and inspect visual heatmaps.

---

Project Structure

├── code/
│   ├── datascience_project.py        # Core PyTorch Autoencoder & Anomalib training script
│   ├── datascience_project_collab.ipynb # Google Colab experiment notebook
│   └── debug.py                       # Diagnostic utilities & testing helpers
├── app.py                             # Gradio web application for visual inspection UI
├── requirements.txt                   # Python environment dependencies
└── .gitignore                         # Configured to exclude heavy binaries & weights

---

Installation & Setup

1. Clone the repository:

   git clone https://github.com/<your-username>/anomaly-detection.git
   cd anomaly-detection
   

2. Create a virtual environment & install dependencies:

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate
   pip install -r requirements.txt
   

---

Model Training

To train the autoencoder model on your own dataset or MVTec AD dataset:

• Run the core script:

  python code/datascience_project.py
  

• Or open code/datascience_project_collab.ipynb in Google Colab for GPU-accelerated training.

---

Running the Gradio Application

To launch the interactive visual inspection UI locally:

python app.py

Open the generated local URL (e.g., http://127.0.0.1:7860) in your web browser.

Note on Model Checkpoint: Place your trained weights (autoencoder_mvtec.pth) in the root project directory before launching app.py.

---

Deploying to Hugging Face Spaces

1. Create a new Space on Hugging Face Spaces selecting Gradio as the SDK.
2. Push app.py, requirements.txt, and your trained model checkpoint (autoencoder_mvtec.pth) to the Space repository.

---

License

This project is open-source and available under the MIT License.