Visual Anomaly Detector

R2026b

Train, evaluate, and compare deep learning anomaly detectors

Since R2026b

expand all in page

Description

The Visual Anomaly Detector app enables you to train, fine-tune, evaluate, and compare deep learning-based anomaly detection models for identifying defects in images.

The app supports Student-Teacher, PatchCore, FastFlow, and FCDD (fully convolutional data description) anomaly detector models. For more information on the characteristics and performance of these anomaly detection models, see Get Started with Anomaly Detection Using Deep Learning.

Using this app, you can:

  • Load and partition labeled image data sets.

  • Train and fine-tune deep learning anomaly detectors using Student-Teacher, PatchCore, FastFlow, or FCDD architectures.

  • Evaluate performance with metrics such as confusion matrices, ROC curves, and precision-recall curves.

  • Optimize detection thresholds using ROC curves and histograms.

  • Visualize predictions with heatmaps showing anomaly locations.

  • Compare multiple trained models to select the optimal detector.

  • Perform inference on production data or live image streams using trained models.

  • Export models, training functions, partitioned datastores, evaluation metrics and plots.

Visual Anomaly Detector app showing trained detectors and performance parameters for trained Student Teacher Detection Model

Open the Visual Anomaly Detector App

  • MATLAB® Toolstrip: On the Apps tab, under Image Processing and Computer Vision, click the app icon.

  • MATLAB command prompt: Enter visualAnomalyDetector. For more information, see Programmatic Use.

Examples

expand all

You can open the Visual Anomaly Detector app with pre-partitioned image datastores loaded from the MATLAB workspace.

Create image datastores for training, calibration, and test sets, and specify the anomaly class labels.

imdsTrain = imageDatastore("data/train",IncludeSubfolders=true,LabelSource="foldernames");
imdsCal = imageDatastore("data/calibration",IncludeSubfolders=true,LabelSource="foldernames");
imdsTest = imageDatastore("data/test",IncludeSubfolders=true,LabelSource="foldernames");
anomalyClasses = ["crack","scratch","contamination"];

Open the app with the preloaded data.

visualAnomalyDetector(imdsTrain,imdsCal,imdsTest,anomalyClasses)

The app opens with the data already loaded and partitioned, so you can immediately select a detector and begin training.

Related Examples

Parameters

expand all

Use the Detector tab to manage sessions, import training data, select and train anomaly detector models, evaluate performance, and export the trained model, training function, data, or results. The parameters and buttons in this tab are usable only after a session has been created or loaded.

Parameters/ButtonsDescription
New Session

Create a new session. Select one of these options from the button menu:

  • Train / Evaluate — Train and evaluate an anomaly detector model by importing training data using the Anomaly Detector: Train / Fine-tune / Evaluate Detector dialog box.

  • Perform Detection — Perform inference with a trained detector on test data using the Anomaly Detector: Perform detection dialog box.

  • Live Detection — Perform inference with a trained detector on a live image stream using the Connect Device dialog box. Using this option requires Image Acquisition Toolbox™.

Open Session

Load a previously saved session MAT file.

Save Session

Save the current session. Select one of these options from the list:

  • Save — Save the current session to the existing MAT file.

  • Save As — Save the current session as a new MAT file.

Visualize

Open the Browse Data tab to browse the imported and partitioned training, calibration, and test data.

StudentTeacher

Select the Student-Teacher anomaly detector and display its configurable parameters in the Hyperparameters panel.

PatchCore

Select the PatchCore anomaly detector and display its configurable parameters in the Hyperparameters panel.

FastFlow

Select the FastFlow anomaly detector and display its configurable parameters in the Hyperparameters panel.

FCDD

Select the FCDD (fully convolutional data description) anomaly detector and display its configurable parameters in the Hyperparameters panel.

Train

Start training the selected anomaly detector model on the imported training data. Training progress is displayed in the Training Progress panel. Click the button again to stop training.

Set Threshold

Open the Threshold panel and configure the anomaly detection threshold.

Confusion Matrix

Display the confusion matrix for the test data predictions at the current threshold.

ROC Curve

Display the receiver operating characteristic (ROC) curve showing the tradeoff between true positive rate and false positive rate across thresholds.

Histogram

Display the histogram of anomaly scores for normal and anomalous images in the test data.

Precision-Recall

Display the precision-recall curve for the test data predictions.

Summary

Display a summary table of evaluation metrics for the trained detector, including accuracy, precision, recall, and F1 score.

Classification Decision

Open the View Decision tab to visualize classification decisions for individual images, including anomaly heatmap overlays.

Compare Trained Detectors

Open the Compare Detectors tab to compare the performance of multiple trained detectors side by side.

Default Layout

Restore the app window to its default panel layout.

Export

Export data to the MATLAB workspace. Select one of these options from the button menu:

  • Detector — Export the trained anomaly detector model.

  • Training Function — Export a MATLAB function that reproduces the training workflow programmatically.

  • Partitioned Datastore — Export the partitioned training, calibration, and test datastores.

  • Performance Metrics — Export evaluation metrics and analysis plots.

Use the Browse Data tab to browse and preview the imported and partitioned image datasets. This tab opens when you select Visualize on the Detector tab.

Parameters/ButtonsDescription
Training Data

Display thumbnail images from the training data set in the image browser.

Calibration Data

Display thumbnail images from the calibration data set in the image browser.

Test Data

Display thumbnail images from the test data set in the image browser.

Thumbnail Size

Adjust the slider to change the size of thumbnail images displayed in the image browser.

Preview

Show or hide a larger preview of the selected image, along with other information, such as assigned label and actual label category.

Close Browse Data

Close the Browse Data tab and return to the Detector tab.

Use the Anomaly Detector: Train / Fine-tune / Evaluate Detector dialog box to import image data and configure a new training session. This dialog box opens when you select New Session > Train / Evaluate on the Detector tab. The dialog box enables you to load data, assign labels, define data distribution, and optionally select pretrained detectors from the workspace.

Parameters/ButtonsDescription
Load data from

Select the source of image data. Choose one of these options:

  • Folder — Select a folder from the file system. The app automatically manages labels from the folder structure.

  • Single Datastore — Import a single ImageDatastore object from the MATLAB workspace.

  • Separate Datastores — Import pre-partitioned training, calibration, and test datastores separately from the MATLAB workspace.

Import ground truth labels

When the selected datastore does not contain labels, you can import ground truth labels separately from the workspace. This option is enabled automatically when labels are not detected in the selected datastore.

Assign labels to data categories

Assign each unique label in the imported data to either the Normal or Anomalous category. Labels assigned to the Anomalous category represent defective images. All other labels are treated as normal.

Workflow type

Select the workflow type:

  • Train & Evaluate — Partition data into training, calibration, and test sets for training a new detector and evaluating its performance.

  • Test — Use all data as a test set to evaluate a pretrained detector imported from the workspace. When you select this option, you must select a trained detector.

Define data ratio

Specify the data split ratios for partitioning the imported data into training, calibration, and test sets. The tree table displays the distribution per label category. You can edit individual values to customize the allocation.

Number/Percent

Choose whether to display the data distribution as absolute image counts or as percentages.

Default

Reset the data split ratios to their default values.

Select detector to import

Select one or more detectors from the MATLAB workspace to import into the session. The table includes all trained anomaly detector objects available in the base workspace. This option is optional for the Train & Evaluate workflow and required for the Test workflow. The app validates that the selected detector image size is compatible with the imported data.

Previous

Return to the previous step.

Next

Proceed to the next step. This button is enabled only after completing the required selections in the current step. On the final step, the button label changes to Done.

Use the Anomaly Detector: Perform detection dialog box to import a pretrained anomaly detector model and test data for performing inference. This dialog box opens when you select New Session > Perform Detection.

Parameters/ButtonsDescription
Import anomaly detector from workspace

Select trained anomaly detector object from the MATLAB workspace. The detector must be a studentTeacherAnomalyDetector, patchCoreAnomalyDetector, fastFlowAnomalyDetector, or fcddAnomalyDetector object.

Import data from

Select the source folder or datastore containing test images for inference.

Import

Load the trained model and test data, and start the inference session.

Use the Hyperparameters panel to configure model architecture and training hyperparameters for the selected anomaly detector. The available parameters change depending on which detector type you select.

Detector Hyperparameters

DetectorParametersDescription
Student-TeacherNetwork

Network size for the Student-Teacher model

Normalization Data Ratio

Ratio of training data for anomaly map normalization. Value ranges from 0.0 to 1.0.

PatchCorePretrained Backbone Network

Backbone network for feature extraction

FastFlowPretrained Encoder Network

Encoder network for feature extraction

Number of downsampling operations

Number of downsampling layers

FCDDPretrained Encoder Network

Encoder network for feature extraction

Number of downsampling operations

Number of downsampling layers

Freeze Backbone

Whether to freeze the backbone network weights during training

For more information on model hyperparameters, see trainStudentTeacherAnomalyDetector, trainPatchCoreAnomalyDetector, trainFastFlowAnomalyDetector, and trainFCDDAnomalyDetector.

Training Hyperparameters

ParametersDescription
Compression Ratio

Ratio for core-set compression. Value ranges from 0.0 to 1.0.

Mini-Batch Size

Number of images processed in each mini-batch during feature extraction

Execution Environment

Hardware for training

Reset Input Normalization

Whether to reset input normalization statistics

Subsampling Method

Strategy for subsampling the core set

Initial Learn Rate

Initial learning rate for the optimizer

Max Epochs

Maximum number of training epochs

Stochastic Solver

Optimization algorithm

Learn Rate Schedule

Learning rate decay schedule

Learn Rate Drop Factor

Factor to reduce the learning rate. Value ranges from 0 to 1.

Learn Rate Drop Period

Number of epochs between learning rate drops

Shuffle

Data shuffling strategy

Momentum

Momentum value for SGDM optimizer

Gradient Decay Factor

Decay rate for first moment estimate (beta1) in Adam optimizer

Squared Gradient Decay Factor

Decay rate for second moment estimate

Epsilon

Small constant for numerical stability

Validation Frequency

Number of iterations between validation evaluations

Validation Patience

Number of validation checks to wait before stopping early

Output Network

Network to return after training

L2 Regularization

L2 regularization factor (weight decay)

Batch Normalization Statistics

Method for computing batch normalization statistics

Gradient Threshold

Threshold for gradient clipping

Gradient Threshold Method

Method for gradient clipping

Execution Environment

Hardware for training

Preprocessing Environment

Data preprocessing parallelization

Acceleration

Automatic acceleration for training

Checkpoint Path

Folder path to save training checkpoints

Checkpoint Frequency

How often to save checkpoints

Checkpoint Frequency Unit

Unit for checkpoint frequency

For more information on these training hyperparameters, see trainingOptions (Deep Learning Toolbox).

Use the Threshold panel to configure and visualize the anomaly detection threshold. This panel opens when you select Set Threshold on the Detector tab. The threshold determines the decision boundary between normal and anomalous predictions.

Parameters/ButtonsDescription
Predefined Threshold

Select one of these threshold method options:

  • Max F1 Score — Maximize the balance between precision and recall.

  • Youdens Index — Maximize the combined sensitivity and specificity.

  • Nearest Zero-One ROC — Select the threshold closest to the ideal point (0,1) on the ROC curve.

  • Custom — Specify a user-defined threshold value.

Threshold Value

Numeric threshold value for classification. This value is automatically set when you select a predefined method. You can enter a custom value when Custom is selected.

Data Selection

Select the data for computing the threshold.

  • Calibration — Use only calibration data for threshold computation.

  • Calibration + Test — Use the combined calibration and test data for threshold computation.

ROC Curve

Display ROC curve with the current threshold position indicated on the curve.

Histogram

Display the score distribution histogram with the threshold line overlaid, showing separation between normal and anomalous scores.

You can adjust the view and export these plots, by clicking the three dots at the top-right corner of the corresponding plots.

Use the View Decision tab to visualize classification decisions for individual images, including anomaly heatmap overlays. This tab opens when you select Classification Decision on the Detector tab.

Parameters/ButtonsDescription
Normalize Anomaly Score

Open a dialog box for configuring anomaly score normalization settings.

Show Heatmap

Toggle the visibility of the anomaly heatmap overlay on images.

Blend Type

Select a blending method for the heatmap overlay. Use "Proportional" to blend the heatmap intensity proportionally with the image, or "Equal" for equal blending.

Colormap

Select the colormap for the anomaly heatmap.

Opacity

Adjust the opacity of the heatmap overlay. Range is 1 to 100.

Thumbnail Size

Adjust the size of image thumbnails in the decision browser.

Preview

Show or hide a larger preview of the selected image with heatmap overlay.

Close View Decision

Close the View Decision tab and return to the Detector tab.

Use the Compare Detectors tab to compare the performance of multiple trained anomaly detectors side by side. This tab opens when you select Compare Trained Detectors on the Detector tab.

Parameters/ButtonsDescription
Model Selection

Select which models to include in the comparison from the left panel which lists all trained detectors. Use Select All to include all detectors or Remove All to clear the selection. The comparison views update dynamically as you change the selection.

Summary

Display a comparison summary table of evaluation metrics for the selected detectors. The table shows metrics such as Global Accuracy, Mean Accuracy, Precision, Recall, Specificity, F1 Score, False Positive Rate, False Negative Rate, and Model Type. Click Select Parameter to customize which metrics appear as columns in the table. Click Export to Workspace to export the summary table to the MATLAB workspace as a table.

Result Plot

Display a bar chart comparing a selected metric across all selected detectors. Use the Select parameter to compare list to choose the metric to plot. Available metrics include Global Accuracy, Mean Accuracy, Precision, Recall, Specificity, F1 Score, False Positive Rate, and False Negative Rate.

ROC Curve

Display overlaid ROC curves for all selected detectors on a single plot, enabling direct comparison of the tradeoff between true positive rate and false positive rate across detectors.

Confusion Matrix

Display confusion matrices for all selected detectors in a tiled layout for side-by-side comparison of classification results.

Precision-Recall

Display overlaid precision-recall curves for all selected detectors on a single plot, enabling direct comparison of the tradeoff between precision and recall across detectors.

Close Compare

Close the Compare Detectors tab and return to the Detector tab.

Use the Connect Device dialog box to connect to an image acquisition device for live anomaly detection. This dialog box opens when you select New Session > Live Detection. This feature requires Image Acquisition Toolbox.

Parameters/ButtonsDescription
Select Device

Select an image acquisition device (camera) for capturing live images. Use the refresh button to refresh the list of available devices.

Select Pixel FormatSelect the pixel format of the image stream.
Connect

Establish a connection to the selected device and begin the live detection session.

Use the Live Detection tab to perform real-time anomaly detection on live image streams captured from a connected image acquisition device. This tab opens when you select New Session > Live Detection and connect a device using the Connect Device dialog box. This feature requires Image Acquisition Toolbox.

Parameters/ButtonsDescription
Detector

Select a trained anomaly detector from the MATLAB workspace. The list includes all anomaly detector objects available in the base workspace. Use the refresh button to refresh the list of available anomaly detectors from the MATLAB workspace.

Start

Begin capturing images from the connected device and running the selected anomaly detector on each captured frame in real time. The button changes to Stop while detection is active.

Heat Map

Toggle the anomaly heatmap overlay on the live image stream. Options are "On" or "Off" (default). This control is enabled only while detection is active.

Opacity

Adjust the opacity of the heatmap overlay on the live image. Range is 1 to 100. Default is 60. This control is enabled only while detection is active.

Switch DeviceAdjust the device properties such as the connected device and pixel format.
Device PropertiesAdjust the properties of the connected device, such as the ROI, frame rate, and saturation.
Live Image Display

View the live image stream from the connected device with real-time classification annotations. The app annotates each frame with the detection result (Good or Bad) and the anomaly score.

Programmatic Use

visualAnomalyDetector opens the Visual Anomaly Detector app.

visualAnomalyDetector(imdsTrain,imdsCal,imdsTest,anomalyClasses) opens the Visual Anomaly Detector app with pre-partitioned image data loaded from the workspace.

  • imdsTrain — Image datastore containing training images, specified as an ImageDatastore object.

  • imdsCal — Image datastore containing calibration images used for threshold optimization, specified as an ImageDatastore object.

  • imdsTest — Image datastore containing test images for evaluating detector performance, specified as an ImageDatastore object.

  • anomalyClasses — Names of the classes that represent anomalous images, specified as a string array or categorical array. The app treats all other class labels in the datastores as normal.

visualAnomalyDetector close closes the Visual Anomaly Detector app.

Version History

Introduced in R2026b

See Also

Functions

Topics