-
Try out the new Jake: AI Coding Assistant for LabVIEW!
Get answers to questions about LabVIEW and discuss your code.
DeepLTK Anomaly Detection Addon by Ngene - Toolkit for LabVIEW Download
Anomaly Detection Addon for DeepLTK
Discussion
| Version | 3.0.1.32 |
| Released | Nov 13, 2025 |
| Publisher | Ngene |
| License | Ngene Custom |
| LabVIEW Version | LabVIEW>=20.0 |
| Operating System | Windows |
| Dependencies | ngene_lv_dnn_deep_learning_toolkit |
| Project links | Homepage Documentation Repository Discussion |
Description
UVAD (Unsupervised Visual Anomaly detection addon for DeepLTK) provides high-performance, annotation-free defect detection for industrial inspection, medical imaging, and predictive maintenance. It learns normal patterns directly from data, enabling accurate detection of unseen anomalies.
Key Features
• High-Speed Training: ~30 sec (GPU), ~5 min (CPU)
• Fast Inference: ~5ms (GPU), ~300 ms (CPU)
• High Accuracy: >95% across 100+ datasets
• Unsupervised Learning: Requires only normal samples, no labels
• Few-Shot Capability: Achieves near-perfect results with a few dozens of good images
• Versatile Applications: Suitable for a wide range of domains including manufacturing, electronics, biomedical imaging, quality inspection, and predictive maintenance.
Examples
Example projects for training, evaluation, and deployment are available on GitHub:
https://github.com/ngenehub/deepltk_examples/tree/main/3_Image_Processing/36_Anomaly_Detection
Tutorials
Step-by-step tutorial covering the full development workflow:
https://www.ngene.co/post/deepltk-tutorial-3-6-visual-anomaly-detection
Release Notes
v3.0.1
This is a major update introducing new functionalities and significant performance improvements.
Note: This version breaks backward compatibility with previous releases of the toolkit.
New Features and Optimizations
1. The inference time was improved by up to 1.75 times due to the latest DeepLTK version (v9.0).
2. Added support for ConvNeXt architecture which performs better (in terms of accuracy and speed) for most of the datasets.
3. Added online Corseting functionality during embeddings generation to speed up the training process and reduce memory requirements for large datasets.
4. Added feature retrieval functionality designed to enable performance analysis.
5. Now NNPC_Predict(Single).vi API returns distance matrix which can be used for feature retrieval functionality.
6. Now Memory Bank (Coreset) generator returns selected features’ coordinates in {Idx, i, j} format to simplify feature retrieval process.
7. Now feature projection layer/s will be skipped if the required projection dimension exceeds the actual feature dimension.
8. Now coreset selection will be skipped if requested Coreset Size exceeds the number of embeddings.
9. Added arbitrary cropping functionality into preprocessing stage.
10. Now, when AUC equals 1, the relative margin is computed and returned instead, allowing comparison between models that achieve 100% accuracy.
11. Optimized NNPC_Eval to skip AUC computation when normal and abnormal distributions do not overlap (AUC = 1).
12. Now NNPC_Eval API sets pixel threshold to the maximum of good pixel score when no pixel ground truth is provided during the training.
13. Optimized evaluation process by efficiently storing ground truth information using DeepLTK's Custom_Data field in NN_Dataset cluster.
14. Added Dataset release functionality into NNPC_Destroy API.
15. Now BIN file can be automatically identified based on CFG file path within NNPC_Init APIs.
16. Added space holder for custom Distance_Metric to Coreset_Params. Custom metrics will be supported in future releases.
17. Restructured and optimized NN output-to-CuLab conversion.
18. Employed dry-run functionality from DeepLTK in NNPC_Init(Train) API to automate support for new layers.
19. Replaced NN_Get_T_dT API with its VIM version from DeepLTK.
20. Strict type defs replaced with Type defs to simplify toolkit upgrade.
21. Other minor improvements.
Bug Fixes
1. Handled warnings in NNPC_Init(Inference) when reading unsupported NNPC parameters from CFG.
2. Fixed selected embeddings index conversion bug in Memory Bank (Coreset) generator.
3. Added validation for selected layers: by checking layer’s dimensionality to be 3D and whether spatial dimensions of selected layers are evenly divisible.
4. Now a warning will be returned if a layer specified by stride is not found in the network.
5. Other bug fixes intended for stability improvements.
Post an Idea
Post a Resource