> For the complete documentation index, see [llms.txt](https://docs.roboflow.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.roboflow.com/models/supported-tasks/anomaly-detection.md).
# Anomaly Detection
Flag defects you have few or no pictures of by training on normal images, then deploy your own anomaly detection model on Roboflow.
## What is anomaly detection?
Anomaly detection learns what a normal image looks like from normal images only, then flags images that differ. For each image, the model returns a "normal" or "anomalous" verdict, an anomaly score, and a heatmap of where the image differs. You do not need pictures of every defect, so it fits problems where defects are rare or vary a lot.
## Supported anomaly detection models
Roboflow supports these anomaly detection models. See the [Anomaly Detection models](/models/supported-models.md#anomaly-detection) table for export and deployment support per model.
## Choosing an anomaly detection model
FoundAD is more accurate, especially when you only have a few normal images. More normal images help both models, and PatchCore the most.
Results on four categories of the VisA dataset (CC BY 4.0).
FoundAD takes the same time per image at any training set size. PatchCore slows down as you add normal images, but at its default 224×224 it stays the faster of the two. See [PatchCore inference speed](/models/supported-models/patchcore.md#inference-speed) for how image size changes this.
PatchCore trains in minutes. FoundAD takes longer, up to hours on large datasets. Try both on your data when you can.
## Fine-tuning an anomaly detection model
The [PatchCore](/models/supported-models/patchcore.md#train-patchcore) and [FoundAD](/models/supported-models/foundad.md#train-foundad) pages cover the split and preprocessing rules for this task.
{% stepper %}
{% step %}
### Create an anomaly detection project
Create a [Project](https://docs.roboflow.com/datasets/create-and-upload/create-a-project) and pick "Anomaly Detection" as the project type.
{% endstep %}
{% step %}
### Upload and label images
[Upload images](https://docs.roboflow.com/datasets/create-and-upload/adding-data), then label each image "normal" or "anomalous" in the [annotation editor](https://docs.roboflow.com/datasets/annotate/annotate/annotation-tools). Most of your images should be normal. The model trains on normal images only. Anomalous images set the decision threshold and measure accuracy, so they go to the valid and test splits only.
{% endstep %}
{% step %}
### Generate a dataset version
[Generate a dataset version](https://docs.roboflow.com/datasets/versions/dataset-versions/create-a-dataset-version) with only the auto-orient and "Stretch to" resize preprocessing steps, and no augmentation.
{% endstep %}
{% step %}
### Train an anomaly detection model
Click "Train" and pick PatchCore or FoundAD. See [Train a Model](/models/train/train-a-model.md) for every option.
{% endstep %}
{% endstepper %}
## Deploy an anomaly detection model
After training, deploy your anomaly detection model with the [Serverless Cloud API](https://docs.roboflow.com/deployment/roboflow-cloud/serverless-api), a [Dedicated Deployment](https://docs.roboflow.com/deployment/roboflow-cloud/dedicated-deployments), or on your own hardware with [Roboflow Inference](https://docs.roboflow.com/deployment/self-hosted/self-hosted). See [Choosing a Deployment](https://docs.roboflow.com/deployment/choosing-a-deployment) to compare the options.
