patchcore-resnet50

Name: patchcore-resnet50
Author: amazon

by amazon

Memory-bank anomaly detection achieving 99.6% AUROC on manufacturing defects

180Kdl/month

25M (ResNet-50 backbone)params

Source Run it on your own data, free

Identifiers

Model ID

amazon/patchcore-resnet50

Feature URI

mixpeek://image_extractor@v1/amazon_patchcore_r50_v1

Overview

PatchCore solves cold-start anomaly detection in industrial manufacturing using only normal (non-defective) images. It builds a maximally representative memory bank of nominal patch-level features from ImageNet-pretrained models, then uses nearest-neighbor outlier detection.

On Mixpeek, PatchCore enables visual quality inspection: upload examples of normal products, and detect defects, anomalies, and deviations automatically.

Architecture

Builds a coreset memory bank of mid-level patch features from a frozen ResNet-50 (ImageNet-pretrained). Uses greedy coreset subsampling for efficient memory. Anomaly scoring via nearest-neighbor distance to the memory bank.

Mixpeek SDK Integration

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors, which returns 13), so patchcore-resnet50 runs
// on your side and the output is upserted through POST
// /v1/namespaces/{namespace_id}/documents/upsert. On Enterprise the other
// path is to upload the weights instead: POST /v1/namespaces/{id}/models
// accepts the huggingface format and a custom plugin loads them.
const res = await fetch(
  "https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer API_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      collection_id: "col_your_collection",
      documents: [
        {
          document_id: "asset-00412",
          // Boxes, masks, depth maps and anomaly scores are structured
          // results, not vectors. They go in payload and are reachable
          // through pre_filters on a retriever, not through similarity.
          payload: {
            detections: modelOutput,
            source_key: "archive/2026/asset-00412",
          },
        },
      ],
    }),
  },
);

// No managed alternative for an open label set. Two extractors do emit a
// bbox, for the one thing each detects: document_graph_extractor@v1 per
// layout block, face_identity_extractor@v1 per face. Nothing ships that
// returns masks, depth maps or anomaly scores.