Title: GPTCLIP: GlobalCLIP Patch Token Framework for Zero-shot Industrial Anomaly Detection and Localization
-Journal/Conference: British Machine Vision Conference (BMVC) 2026
-Authors: Junha Park, Jongpil Jeong
-DOI: 
-Journal/Conference Link: https://bmvc2026.org/

Abstract: In smart manufacturing environments, the key challenge of zero-shot anomaly detection (ZSAD) is to detect defects without re-learning new categories. However, existing
CLIP-based ZSAD methods suffer from two key limitations: dependence on text prompt
engineering or auxiliary normal samples, and uniform processing of multi-layer patch
features that ignores per-layer discriminative differences. To address both, this paper
proposes a two-stage framework that uses neither text prompts nor auxiliary data. Stage 1
computes an image-level anomaly score s1 via a GlobalCLIP-based autoencoder ensemble, and Stage 2 processes seven intermediate layer patch tokens of CLIP ViT-L/14@336
via Position-aware k-NN to generate a pixel-level anomaly map. Two key modules reinforce the framework: SIG (Score-guided Integration) dynamically corrects the pixel
map through s1 to maximize the synergy between the two stages, and DALW (DiversityAware Layer Weighting) automatically selects the optimal layer combination per category through log-scale stability and diversity-based adaptive weights. On MVTec AD,
we achieve P-AUROC 97.59% and AUPRO 90.98%, surpassing all methods that rely on
text prompts or auxiliary data by 2.39%p and 8.18%p, respectively, while remaining on
par in P-AUROC with the strong text- and data-free transductive baseline MuSc (97.3%).
On VisA, we achieve competitive P-AUROC 88.74%.

-Status: Submitted (2026.05.29)