all projects
Inspection

ViT-AnomalyDetection

A Vision-Transformer autoencoder for industrial inspection: reconstruct the image, then localize defects from patch-wise reconstruction error.

Result Pretrained ViT encoder + custom decoder for unsupervised anomaly localization Stars ★ 5 Tags ViT · Autoencoder · Anomaly Detection · Inspection
View on GitHub

An unsupervised industrial-inspection model: train only on “good” parts, then flag anything the network cannot reconstruct well.

Approach

  • ViT autoencoder. A pretrained Vision-Transformer encoder feeds a custom decoder that reconstructs the input image.
  • Patch-wise error → localization. Anomalies are found where reconstruction error is high, producing a per-patch heatmap rather than just an image-level score.
  • Why it matters. This maps directly to the kind of defect-detection work I did as a Computer Vision Engineer at Schwarz IT.