🔍 Anomaly Detection

Anomaly-based defect detection system for industrial quality control, built with PatchCore and Streamlit.

Part of the Building AI-Powered Defect Detection Systems for Industrial Quality Control course.

Machine Vision Build Project Steel Defect Classifier

By Zuriel Olu-Silas

This project implements a custom Convolutional Neural Network (CNN) to detect and classify surface defects in steel images using the Severstal Steel Defect Detection Dataset.

Title: Steel Defect Classifier DatasetSource: Severstal Steel Defect Detection DatasetClasses: no_defect, defect_1, defect_2, defect_3, defect_4 Total images: 1,000 (Balanced subset of 200 images per class)

Dataset

Class	Count
no_defect	200
defect_1	200
defect_2	200
defect_3	200
defect_4	200

Train / Test Split

The dataset was split using a stratified approach to maintain class balance across all sets.

Split	Size
Train	699 (70%)
Validation	150 (15%)
Test	151 (15%)

Model Architecture

The SteelCNN is a custom convolutional neural network designed for industrial surface inspection. It uses batch normalization for training stability and dropout in the fully connected layers to prevent overfitting.

SteelCNN(
  (features): Sequential(
    (0): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (1): BatchNorm2d(32)
    (2): ReLU()
    (3): MaxPool2d(kernel_size=2, stride=2)
    (4): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (5): BatchNorm2d(64)
    (6): ReLU()
    (7): MaxPool2d(kernel_size=2, stride=2)
    (8): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (9): BatchNorm2d(128)
    (10): ReLU()
    (11): MaxPool2d(kernel_size=2, stride=2)
  )
  (pool): AdaptiveAvgPool2d(output_size=1)
  (classifier): Sequential(
    (0): Flatten()
    (1): Linear(in_features=128, out_features=64)
    (2): ReLU()
    (3): Dropout(p=0.3)
    (4): Linear(in_features=64, out_features=5)
  )
)

Trainable parameters: 102,277

Training Results

The model was trained for 10 epochs using the Adam optimizer (with a learning rate of 0.001) and Cross-Entropy Loss.

• Best Validation Accuracy: 0.513 (Reached at Epoch 8)
• Final Training Loss: 1.2607
• Inference Latency: ~3ms per frame (on NVIDIA T4 GPU)

Training Loss Curve

Evaluation - Confusion Matrix

The model demonstrates an ability to distinguish between defect-free surfaces and specific defect types, though there is some overlap between similar defect classes.

Quick Start

# Install PyTorch first (see docs for CUDA/MPS options)
pip install torch torchvision

# Install project dependencies
pip install -r requirements.txt

# Train PatchCore on MVTec Metal Nut (one-time, ~2 min)
python -m anomaly_detection.train

# Launch the inspection app
streamlit run anomaly_detection/app.py

Project Structure

anomaly_detection/
├── anomaly_detection/     # Main Python package
│   ├── app.py             # Streamlit frontend
│   ├── acquisition.py     # Camera simulator
│   ├── inference.py       # PatchCore model inference
│   ├── preprocessing.py   # Image transforms
│   ├── train.py           # Model training script
│   └── utils.py           # Config, logging, paths
├── .streamlit/            # Streamlit theme config
├── data/                  # MVTec dataset (gitignored)
├── models/                # Saved model checkpoints
├── docs/                  # MkDocs documentation source
├── tests/                 # Unit tests
├── .github/workflows/     # CI/CD pipeline
├── Dockerfile             # Container deployment
├── requirements.txt
├── Zuriel_Steel_Defect_Project (2) # MY ACTUAL WORK IS HERE
└── mkdocs.yml             # Docs configuration

Documentation

mkdocs serve    # Preview at http://localhost:8000

License

This project is for educational purposes as part of the Building AI-Powered Defect Detection Systems for Industrial Quality Control course.