Fabric Surface Anomaly Detection

Overview

Fabric inspection requires detecting subtle defects — snags, holes, density variations — with no defective samples available during training. This project adapts the PatchCore algorithm for high-resolution textile images.

Method

Patch features are extracted from WideResNet-50 at layers $\ell_2$ and $\ell_3$, then aggregated into a memory bank $\mathcal{M}$:

The anomaly score for a full image is the maximum patch score:

Coreset Subsampling

The memory bank is compressed via greedy coreset selection to maintain $|\mathcal{M}| \leq 1\%$ of the original feature set with < 1% AUROC degradation. This reduces search time with FAISS from 120 ms to < 5 ms.

Results

| Metric | Value | |---|---| | AUROC | 98.9% | | Inference Time | 12 ms/frame | | Training Samples Needed | 200 (normal only) |