---
license: other
tags:
  - hyperspectral
  - anomaly-detection
  - dinomaly
  - cuvis-ai
library_name: cuvis-ai
---

# Dinomaly - bedding (6-channel VIS+SWIR) anomaly detector

Trained Dinomaly (Anomalib) pipeline for foreign-object anomaly detection on the
bedding hyperspectral dataset
([cubert-gmbh/X4_SWIR_Industrial_Foreign_Object_Detection_Bedding](https://huggingface.co/datasets/cubert-gmbh/X4_SWIR_Industrial_Foreign_Object_Detection_Bedding)).

- 6-band selector (625, 550, 450, 1450, 1200, 1050) nm, input_channels=6
  (duplicate-and-halve patch-embed inflation), image_size=672 (square squash).
- Trained 20 epochs, fp32, on the 1800x4300 center crop of the native 2400x4900 cubes.

## Files
- dinomaly_bedding_all6.yaml / .pt - serialized cuvis-ai pipeline (load with CuvisPipeline.load_pipeline).
- eval_val/ - validation metrics (report.json, per-class AUROC, Dice).

## Headline (val, 59 frames)
pixel AUROC 0.976 - image AUROC (filename) 0.984 - Dice@F1 0.615 - mean per-class AUROC ~0.95.

Used by the bedding tutorial notebooks in cuvis-ai-dinomaly/notebooks/bedding_anomaly/.
Requires the cuvis SDK + high-level cuvis-ai to load (pipeline uses cuvis_ai.node.* built-ins).