Unsupervised anomaly detection system for industrial inspection using PatchCore + ResNet50.
Learns normal patterns only and detects defects as deviations, making it suitable for real-world factory environments with unknown or rare failures.
- No defect labels required
- Detects unseen anomalies
- PatchCore memory bank inference
- GPU support (CUDA, FP16)
- ONNX / TensorRT export
- Continuous feedback loop for improvement
Train on normal images -> build memory bank -> score new samples -> flag anomalies -> collect edge cases -> retrain.
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt
bash scripts/full_pipeline.sh screw data/raw/screw/test- PatchCore (anomaly detection)
- PyTorch + CUDA (GPU acceleration)
- MVTec AD dataset
- ONNX / TensorRT (deployment)
The system learns what normal looks like and flags anything that deviates as anomalous.
Industrial inspection
Surface defect detection
PCB / metal part quality control
A practical anomaly detection pipeline designed for industrial environments with unknown and evolving defects.