R&D
CENTER
KOENJA

Choosing an Anomaly Detection Algorithm: 11 Options and When to Use Each

The 11 algorithms in AD Workbench, including PaDiM, PatchCore, AnomalyDINO, WinCLIP and EfficientAD, produce the same outputs but have different strengths. This lesson gives criteria for choosing by data volume, part fixturing, resolution and takt time.

Core4 min0 views
Key points
  1. Algorithms fall into two groups: collect-in-one-pass methods (1-pass) and neural-network training methods (gradient descent).
  2. With fewer normal images, pretrained models (WinCLIP, AnomalyDINO) work better. With more, PatchCore and Dinomaly2 work better.
  3. For parts held in a jig, PaDiM and PatchCore (position-specific reference) fit well. For parts whose position varies, AnomalyDINO is stronger.
  4. If the choice is hard, check several algorithms and compare them on the same data in one go with Auto Run. This is the most reliable way.

All AD Workbench algorithms produce the same outputs: a per-pixel anomaly heatmap and one score for the image. The decision is score ≥ Threshold → NG. They differ in what they use to extract features and how they measure anomaly. That difference sets the data volume they need, their speed and the defect types they handle best.

Two Training Approaches

A. Statistics / feature collection (1-pass) — PaDiM, PatchCore, SPADE, AnomalyDINO, WinCLIP, HoleMetric, two Tiled variants

These scan the normal images once, collecting features or computing mean and covariance. There is no backpropagation, so there are no epochs. Training takes a few seconds to a few minutes. They store the collected features, so the model file can be large.

B. Gradient descent training — EfficientAD, FastFlow (20 epochs), Dinomaly2 (5,000 iterations)

These actually train network weights, so they take longer. In return, inference is fast and the model file is small. They are candidates when the line takt time is tight.

The 11 at a Glance

Algorithm Feature extraction Anomaly measure Main strength
PaDiM WideResNet50-2 Mahalanobis distance to a per-position Gaussian Fixed-position targets, sharp defect location
PatchCore WideResNet50-2 k-nearest distance to a memory bank Top-tier precision (recommended default)
SPADE WideResNet50-2 Per-position minimum distance to K similar normal images Simple, fast setup with small data
WinCLIP CLIP ViT-B-16+ Similarity to "normal/defective" text Try immediately with almost no normal images
EfficientAD Lightweight dedicated network Teacher–Student mismatch + reconstruction error Very fast inference
FastFlow WideResNet50-2 Probability density of a normalizing flow Small model, fast inference
AnomalyDINO DINOv2 ViT Nearest cosine distance of patch tokens Complex shapes, few-shot, position variation
Dinomaly2 DINOv2 + decoder Encoder–decoder feature difference Merges several views into one model
HoleMetric Classical vision Pixel/cell statistics after alignment Fine hole and cell shapes, CPU, 3 s training
AnomalyDINO Tiled DINOv2 ViT-B/14 Per-tile banks at native resolution Fine defects at high resolution, foreign material, cracks
PatchCore Tiled WideResNet50-2 Per-tile coreset banks 20+ normal images, fine surface defects

What Foundation Models Changed

WinCLIP, AnomalyDINO and Dinomaly2 use the features of general-purpose vision models pretrained on more than 100 million images (CLIP, DINOv2). These models already represent object structure well. Even with only a few normal images, they extract features that separate normal from anomalous. WinCLIP uses CLIP, which learned images and text together. Even with zero normal images, it tries to judge by similarity to the text "part without flaws" and "scratched part" (zero-shot). It is useful for judging whether inspection is feasible before you collect data.

Fine Defects at High Resolution: Tiling

Most models shrink the image to around 512 pixels. If you shrink a 2048-pixel original to a quarter, a 4-pixel-diameter blocked hole becomes 1 pixel and disappears. The tiled approach first aligns the part. It then splits the image at native resolution into 518-pixel tiles (10% overlap) and builds a separate bank for each tile. One hole appears as a 4×4 patch in the original, so cell-level defects stay in the features.

HoleMetric uses alignment and statistics, with no deep learning. It measures brightness, dark ratio and deviation for each hole (cell). It explains which cell caused an NG in cell_report.csv. We recommend using it together with AnomalyDINO Tiled, which is strong on irregular anomalies (foreign material, cracks).

Selection Guide

Situation Recommended Why
0–10 normal images (feasibility check) WinCLIP · AnomalyDINO Try right away with zero/few-shot
10–50 normal images AnomalyDINO · SPADE · HoleMetric · Tiled Training-free bank or statistics methods
50–300 normal images PatchCore · AnomalyDINO · PaDiM Top-tier precision
300+ normal images PatchCore · Dinomaly2 Dinomaly2 covers several views with one model
Fixed in a jig, always the same spot PaDiM · PatchCore · HoleMetric Position-specific reference matches exactly
Target moves or rotates slightly AnomalyDINO Relatively robust to position change
Complex shapes such as connectors and molded parts AnomalyDINO · Dinomaly2 Better at representing complex structure
Fine hole and pin shapes (2048px) HoleMetric + AnomalyDINO Tiled Cell-level measurement at native resolution
Tight inline takt time EfficientAD · FastFlow Fast inference, light model

In the End, Decide by Experiment

The table above is only a starting point. For the same part, the ranking changes with lighting and defect type. In AD Workbench, check several candidate algorithms in Setup and press Auto Run. It runs all of them on the same training and evaluation data, so you can compare score distributions and heatmaps side by side. Model Evaluation and Field Operations covers the comparison criteria.