phenotypic.util.medoid_ciede2000

phenotypic.util.medoid_ciede2000(lab_points: numpy.ndarray, max_pixels: int = 1000, seed: int = 0, chunk_size: int = 128) → tuple[TypeAliasForwardRef('numpy.ndarray'), TypeAliasForwardRef('numpy.ndarray')]

ΔE2000 medoid center and per-pixel ΔE2000 distances to it.

The medoid (real pixel minimizing total ΔE2000) is selected from a seeded subsample of at most max_pixels; the returned distances are computed from the chosen medoid to all input pixels.

The total-distance (“row sum”) used to pick the medoid is accumulated in candidate blocks of chunk_size rows rather than materializing the full m x m pairwise matrix. Peak memory is therefore O(chunk_size * m) instead of O(m^2) (CIEDE2000 allocates many intermediate arrays the size of its broadcast grid). The result is bit-identical to the full-matrix form for any chunk_size – chunking bounds memory, not accuracy.

Parameters:

• lab_points (numpy.ndarray) – (N, 3) CIE L*a*b* pixel vectors.

• max_pixels (int) – Subsample cap for medoid selection.

• seed (int) – RNG seed for reproducible subsampling.

• chunk_size (int) – Number of candidate rows scored per block; caps peak memory.

Returns:

(center (3,), all_deltas (N,)). center is all-NaN and all_deltas empty when lab_points is empty.

Return type:

tuple[TypeAliasForwardRef(‘numpy.ndarray’), TypeAliasForwardRef(‘numpy.ndarray’)]