Few-Shot Anomaly Detection

When a new product enters the line, there may be only a handful of normal images available, sometimes only one. Few-shot anomaly detection works from exactly that: a tiny set of normal "support" images that defines what normal looks like.

Which images end up in that set turns out to matter a great deal, and it matters more once a single model has to serve many categories at once. A support image that happens to be unrepresentative, or that sits too close to a different category in feature space, quietly distorts the model's notion of normality. There is a second, separate failure: patch-based detectors score a whole image by pooling its highest-scoring patches, so a strong response from a shadow, a background texture, or a fixture, anywhere outside the product itself, can dominate the final decision.

Evaluated on MVTec-AD, VisA, and Real-IAD, DDR performs consistently well, with the clearest margins on the large-scale Real-IAD benchmark, where it surpasses the Dinomaly baseline on every image-level and pixel-level metric. Ablations show a two-stage pattern: CLC delivers the dominant stabilization gain, while DLA and DSR contribute a complementary gain that only appears once the target has been calibrated.

The same model transfers to medical imaging without modification. On the BMAD benchmark it reaches 93.92 image-level AUROC on brain MRI and 93.94 on retinal OCT, ahead of the memory-bank method PatchCore (91.65 and 91.55). Throughout, DDR stays simple, purely reconstruction-based, and entirely free of synthetic-anomaly supervision.