Detect Image Anomaly (PaDiM)

@cealist1 VersionDetect

PaDiM anomaly detector: takes an Image and emits {score: Double, heatmap: Image} — an image-level anomaly score plus a per-pixel heatmap Image. Use when a class-specific feature distribution is fitted ahead of time and per-pixel localisation is needed.

How it fits

Typical backends

Live visual QA: each frame's anomaly heatmap is drawn back over the part for an operator to watch defects light up in real time.

Automated reject gate: the anomaly score is thresholded by an expression stage and only offending parts are reported as JSON to a line controller.

Caveats

Limitationmodel_path and device are fixed at startup; swapping the checkpoint or backend requires a redeploy.
FallbackRequesting a GPU when none is available falls back to CPU at startup, and an empty input Image short-circuits to a zero score and a 1x1 black heatmap without running the network.
AccuracyThe heatmap is min-max normalised per frame, so absolute heatmap intensities are not comparable across frames; compare the scalar score instead.
Hard constraintPaDiM fits a per-class Gaussian over backbone features, so the checkpoint misbehaves on any product class it was not fitted on — there is no automatic out-of-distribution rejection.

Versions

86faaa72latestdefault
Automated release
6/29/2026