ML Project

Overview

A training-free anomaly detection system for industrial inspection built on the PatchCore methodology. The system constructs a memory bank from normal images during a one-time setup phase and detects anomalies at inference time through nearest-neighbor distance calculations — no gradient updates or retraining required.

WideResNet-101-2 is used as the backbone for multi-scale patch descriptor extraction with L2 normalization. Greedy coreset subsampling retains approximately 1% of patch vectors to keep the memory bank compact and efficient. FAISS-GPU enables fast KNN lookups, while Gaussian smoothing refines the resulting anomaly heatmaps.

Evaluated on the MVTec Anomaly Detection benchmark across 15 industrial object categories. Achieved 97.07% mean image-level AUROC and 95.88% pixel-level AUROC, with per-image inference running in 5–15 ms on an H100 GPU.