Reproduce “ReConPatch : Contrastive Patch Representation Learning for Industrial Anomaly Detection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jeeho Hyun, Sangyun Kim, Giyoung Jeon, Seung Hwan Kim, Kyunghoon Bae, Byung Jun Kang

ReConPatch is an unsupervised metric learning framework for industrial anomaly detection that generates discriminative features by linearly modulating pre-trained patch features using contrastive learning with pseudo-labels derived from pairwise and contextual similarities. The method achieves state-of-the-art anomaly detection performance on MVTec AD (99.72% image-level AUROC with ensemble) and BTAD datasets while reducing the need for extensive data augmentation.

Key insights

  • Combining both pairwise and contextual similarities effectively generates robust pseudo-labels for unsupervised contrastive learning, enabling the arrangement of feature space for normal data without explicit anomaly labels.
  • A lightweight linear modulation of pre-trained features can sufficiently mitigate the distribution shift between natural and industrial images, enhancing feature discriminability for industrial anomaly detection.
  • The training process implicitly encourages features corresponding to similar spatial positions to group more closely, suggesting that ReConPatch learns positional information and generates more structured feature spaces.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.