Patch-Level DINOv2 Scoring Overcomes Signal Dilution in LIGO Glitch Detection

This arXiv preprint (v1, June 2026) introduces an unsupervised patch-level architecture using frozen DINOv2 (ViT-S/14) to detect gravitational-wave glitches in spectrograms. Unlike prior global CLS-token methods that average over 1369 patches and suppress signals under 5% occupancy, the approach substitutes top-k order statistics on individual patch tokens against a vector-quantized index of 1216 centroids (K=64 per class across 19 Gravity Spy O3b morphologies). Tested on strain-domain injections in a single LIGO O4a L1 session (20260524), it yields strong distributional separation (KS=0.963 at k=68) for extended morphologies like SpiralBurst while highlighting resolution limits for ultra-short events such as AsymBlip. A topological saliency map derived from 78 null segments enables localization without binary classification. As a preprint, results remain unpeer-reviewed; the non-isotropic embedding geometry noted by the authors suggests saliency maps function more as visualizers than detectors. This builds on Gravity Spy’s citizen-science morphology catalog (Zevin et al. 2017) and earlier CNN glitch classifiers (e.g., Cuoco et al. 2022), addressing a gap mainstream coverage overlooks: preserving faint astrophysical signals during real-time vetoing. Limitations include single-session data and patch-size temporal constraints.