Skip to content

Few-shot learning

Few-shot is not magic: launching inspection with limited examples

The field rarely starts with thousands of labelled defects. Normal examples are common; critical defects are scarce and products keep changing.

About 1 min read

Back to the journal

Few-shot anomaly detection can lower startup cost, but representative references, consistent imaging and explicit acceptance boundaries still determine the outcome.

Start from normal references and a few examples

Many teams cannot assemble a large defect dataset before a project begins. Few-shot methods instead learn from a small number of normal references or transfer visual-language knowledge to a new object. This makes an incremental launch possible, but it does not remove the need for reliable imaging and representative variation.

Use research to choose a validation path

PatchCore demonstrates the value of a representative memory of local visual features, while WinCLIP explores zero- and few-shot anomaly classification and segmentation with vision-language representations. The product lesson is to start with stable inspection items, collect edge cases and accept capability one item at a time rather than promise universal coverage.

Validate new experience before release

Learning from reviewed samples must be controlled. Candidate experience should carry its source, version, comparison metrics and rollback point. An inspector should see what changed and which items are affected without needing to understand feature banks or prompt tuning.

Academic sources

  1. Towards Total Recall in Industrial Anomaly DetectionCVPR 2022 / arXiv

    A representative local-feature memory approach for industrial anomaly detection.

  2. WinCLIP: Zero-/Few-Shot Anomaly Classification and SegmentationCVPR 2023 / arXiv

    Vision-language representations for zero- and few-shot anomaly classification and segmentation.