anomaly_patch op• Data kinds: video → table
• Call: import fullseye as fs; fs.ledger.patchdist_fit(stack, patch=8, stride=4, scales=(1.0, 2.0, 4.0), groups=('gradient', 'gabor', 'lbp'), n_orient=4, shrinkage='auto', n_dims=None, seed=0, ridge=1e-06, pool='position') (to call the implementation directly, import anomaly_patch; anomaly_patch.patchdist_fit(stack, patch=8, stride=4, scales=(1.0, 2.0, 4.0), groups=('gradient', 'gabor', 'lbp'), n_orient=4, shrinkage='auto', n_dims=None, seed=0, ridge=1e-06, pool='position'); from the registry, opsspc.get("patchdist_fit"))
Learn a normal distribution of the patch feature at every grid position from good images (table).
stack is (n, H, W) or (n, H, W, 3) of good images, n >= 2, registered well enough that a grid cell sees the same content. The feature is patch_features (same parameters). At every cell the mean and the covariance (maximum likelihood, divided by n) of the D-dimensional feature are estimated; n <= D makes the covariance singular, so the shrinkage toward tr(S)/D * I is declared: "auto" is the Ledoit–Wolf analytic intensity per position, a number in (0, 1] fixes it, "none" (or 0) is accepted only when n - 1 >= D and the covariance is positive definite and otherwise refused. n_dims keeps a seeded random subset of the D dimensions (the memory trick of the original method), ridge (relative to the median variance, applied unless shrinkage="none") is a numerical floor against exactly constant channels. pool="position" (default) estimates one distribution per grid cell, right when the content is registered; pool="global" estimates ONE distribution from the patches of all cells and images (n * gh * gw samples, overlapping patches are not independent), right for a stationary texture whose content at a cell differs from image to image and which has far more samples than dimensions.
Returns the model: per-position mean and precision (inverse of the shrunk covariance), shrinkage_map and its mean shrinkage_mean, n_train, dim, dims (kept channels), the feature parameters and image_shape. Pass it to patchdist_score.
Raises ValueError for fewer than two images, a singular covariance with shrinkage="none", a constant feature that stays singular, or invalid parameters.
• Sample-data catalog (download URLs / licences) — 2-D uses skimage.data (BSD/public domain) plus synthetic images; 3-D lists download URLs for real data sources (Stanford, PDS, …).
• Operator provenance and references — the sources of the research/methods this op family came from.
• The canonical algorithm (author, year) and its uses are named in the family usage guide above.
• poc_patch_anomaly — py -3.11 examples/poc_patch_anomaly.py
table as input)msa_anova_table · msa_gauge_rr · msa_bias_linearity · msa_attribute_agreement · gum_standard_uncertainty · gum_propagate · gum_expanded · gum_monte_carlo
anomaly_patch)patch_features · patchdist_score · patchbank_fit · patchbank_score
*Provenance: anomaly_patch.py — SPC operator registry. This per-op note is generated by tools/opdocs.py md (do not hand-edit).*
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.