Feature Extraction

A “feature map” is a new image, the same size as the original, where each pixel’s value describes something about that spot — like how sharp an edge is there, or how much texture surrounds it — instead of its raw brightness. Feature extraction is a built-in capability of SPICEData, reached through its features property. spice.features.extract() computes several such feature maps from spice.data — Gaussian, Sobel, Gabor, entropy, Local Binary Pattern (LBP), and more — and returns them as a new SPICEData, sharing the same history as the object you called it on.

spice.features

No separate function call with an explicit spice argument is needed — spice.features is accessed directly on any loaded SPICEData:

from juspice.io import load_data

spice = load_data('image.tif')
feature_spice = spice.features.extract()

Unlike juspice.preprocess_module.PreprocessingAccessor, which mutates spice.data in place and returns spice itself, extract() returns a derived object — the feature stack has a different shape than the source image — so the original spice is left untouched:

feature_spice = spice.features.extract(feature_names='sobel_magnitude')

feature_spice is spice            # False — a new SPICEData
feature_spice.history is spice.history   # True — shared lineage

The call is recorded automatically into spice.history using the actual runtime argument values — no manual tracking calls are needed.

Methods

  • extract() — compute a feature-map stack and return it as a new SPICEData. Accepts feature_names=None (all features), a single feature name, or a list of names.

  • maps() — read-only: return all feature maps as a dict (not recorded into history).

  • stack() — read-only: return the full feature-map stack as an (H, W, N) array.

  • dict() — read-only: scalar summaries (mean of each feature map).

  • vector() — read-only: scalar summary vector in canonical FEATURE_MAP_NAMES order.

Available feature channels

  • Gaussian smoothed at multiple scales,

  • Sobel, Scharr, and Laplace edge/gradient response,

  • Hessian, Frangi, Meijering, and Sato ridge/vesselness response,

  • Gabor filter bank responses,

  • Local Binary Patterns (LBP),

  • Canny edges,

  • Shannon entropy (sliding window),

  • Structure-tensor coherence.

See FEATURE_MAP_NAMES for the full, ordered list of 16 feature names.

Saving feature results

Use save_feature_results() to write the feature array and its provenance:

from juspice.feature_extract import save_feature_results

save_feature_results(feature_spice)

This calls save_data() internally and writes <stem>.npy, <stem>.json, and <stem>_history.py to the auto-detected output location.

Reproducibility

Operations applied through spice.features are recorded into DatasetHistory as features.extract entries, using the actual runtime argument values. They appear in the <stem>.json sidecar and reconstruct as feature_spice = spice.features.extract(...) in to_lines() and to_script().

When used inside a Tracker recording block, every call is additionally captured verbatim in the session-level _history.py.

See also